{"branches":[{"author":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-05-20T14:39:26.000000Z"},"committer":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-05-20T14:39:26.000000Z"},"message":"Add Discord Todoist action handling\n","name":"codex/discord-todoist-actions","oid":"25f1e8767741532cad7b9e8b8050c31b916cc384"},{"author":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-05-15T01:58:11.000000Z"},"committer":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-05-15T01:58:11.000000Z"},"message":"Add Discord model provider slash option\n","name":"codex/discord-model-provider","oid":"e6f509ba3595c0a0568327c1f7f2ea74e7569305"},{"author":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-04-30T22:00:32.000000Z"},"committer":{"email":"aliceisjustplaying@gmail.com","name":"alice","when":"2026-04-30T22:00:32.000000Z"},"message":"fix(cli): preserve compact response padding after update\n","name":"main","oid":"56a916e6c474f63b6664dcd81f01fd24bb06bcda"},{"author":{"email":"pickett.austin@gmail.com","name":"Austin Pickett","when":"2026-04-30T17:41:10.000000-04:00"},"committer":{"email":"pickett.austin@gmail.com","name":"Austin Pickett","when":"2026-04-30T17:41:10.000000-04:00"},"message":"feat(dashboard): add Plugins page with enable/disable, auth status, install/remove\n\n- New PluginsPage.tsx: full plugin management UI (list, enable/disable,\n install from git, remove, git pull updates, provider picker)\n- Backend: dashboard_set_agent_plugin_enabled now also toggles the\n plugin's toolset in platform_toolsets so enabling actually makes\n tools visible in agent sessions\n- Backend: /api/dashboard/plugins/hub returns auth_required + auth_command\n per plugin (checks tool registry check_fn)\n- Frontend: auth_required shown as Badge + CommandBlock with copy-able\n auth command\n- Fix: Select overflow in providers card (min-w-0 grid cells, removed\n truncate/overflow-hidden that clipped dropdown)\n- Refactor: _install_plugin_core extracted for non-interactive reuse,\n PluginOperationError for structured error handling\n- i18n: en/zh/types updated with all new plugin page strings\n","name":"austin/feat/plugins-page","oid":"7947b8971096b393ba10049882be55a086d39f49"},{"author":{"email":"arilotter@gmail.com","name":"ethernet","when":"2026-04-30T13:49:38.000000-04:00"},"committer":{"email":"arilotter@gmail.com","name":"ethernet","when":"2026-04-30T17:26:22.000000-04:00"},"message":"feat(tui): add ability to stash and pop prompts\n\nwith ctrl-s / ctrl-p, letting you interrupt your own prompt to do\nsomething else & bring it back later\n","name":"feat/stash-prompt","oid":"b7f90e5cef64bb84bc708c990cfeb02f6eaa8b68"},{"author":{"email":"41898282+github-actions[bot]@users.noreply.github.com","name":"github-actions[bot]","when":"2026-04-30T19:49:01.000000Z"},"committer":{"email":"41898282+github-actions[bot]@users.noreply.github.com","name":"github-actions[bot]","when":"2026-04-30T19:49:01.000000Z"},"message":"fix(nix): refresh npm lockfile hashes\n","name":"fix/bundle-size","oid":"279504d5b887ad57a7da4c3ca8d7d534d7ca47fe"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-30T12:36:02.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-30T12:36:02.000000-05:00"},"message":"fix(docker): run bundled TUI without npm install\n","name":"bb/docker-tui-prebuilt","oid":"d049d88dd7f53643749b0ce91bc65e8f78dd0f87"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"Teknium","when":"2026-04-30T10:19:08.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"Teknium","when":"2026-04-30T10:19:08.000000-07:00"},"message":"test(auxiliary): guard raw-URL plumbing for Anthropic-compat endpoints\n\nAdds a regression suite for PR #17467 (issues #17705, #17413, #17086,\n#10469). Two layers of guards:\n\n1. Primitive-level: `_endpoint_speaks_anthropic_messages` and\n `_maybe_wrap_anthropic` behave correctly on both raw `/anthropic`\n and rewritten `/v1` inputs — so the detector/wrapper pair cannot\n silently drift.\n2. Call-site plumbing: `_resolve_api_key_provider()` exercised end-to-end\n with a simulated MiniMax pool entry and a simulated explicit-creds\n entry. Both must forward the RAW `/anthropic` URL to\n `_maybe_wrap_anthropic`. Reverting any of the 4 `raw_base_url`\n plumbing changes in #17467 produces a clear failure:\n `Got: https://api.minimax.chat/v1, expected /anthropic`.\n\nValidated by planting the pre-fix code: 2/2 call-site tests fail with\ndiagnostic messages; primitive tests correctly remain passing (they\nwere always true at the helper level).\n","name":"fix/aux-anthropic-url-regression-test","oid":"e7d84348d321d5afa6e9d24d19ed1409757784e3"},{"author":{"email":"41898282+github-actions[bot]@users.noreply.github.com","name":"github-actions[bot]","when":"2026-04-30T15:45:16.000000Z"},"committer":{"email":"41898282+github-actions[bot]@users.noreply.github.com","name":"github-actions[bot]","when":"2026-04-30T15:45:16.000000Z"},"message":"fix(nix): refresh npm lockfile hashes\n","name":"dont-mergeme/test-fail-ci","oid":"847cf0335a853fc2e10fc2bb8b4b937d3ad3f963"},{"author":{"email":"vincent.brunet.us@gmail.com","name":"crypto-vincent","when":"2026-04-30T18:52:53.000000+09:00"},"committer":{"email":"vincent.brunet.us@gmail.com","name":"crypto-vincent","when":"2026-04-30T18:52:53.000000+09:00"},"message":"add-to-release-contributor-map\n","name":"vbrunet/2026_04_15-terminal-title-updates","oid":"64d9755a79b041e69a029c7305c8a272a3942185"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-29T19:23:31.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-29T19:23:31.000000-05:00"},"message":"feat(skins): add bunnny — barbie-pink coquette theme ♡\n\nAdds a built-in 'bunnny' skin preset with a hot-pink coquette palette:\n\n- Hot pink (#FF3366) borders with Barbie-pink (#FF69B4) accents\n- Lavender-blush (#FFF0F5) text on deep-plum (#2A0E1E) surfaces\n- Coquette spinner verbs (sparkling, twirling, tying a little bow)\n- Heart/sparkle/flower spinner faces (♡ ✧ ✿ ❀ ෆ)\n- Heart (♡) prompt symbol and tool prefix\n- (ﾉ◕ヮ◕)ﾉ*:･ﾟ✧ kaomoji in welcome + help header\n- Custom HERMES <3 banner_logo in pink gradient\n- banner_hero of twin coquette bunnies holding paws, framed with\n floating sparkles, hearts, and flowers to fill the banner width\n\nSkin is cosmetic only — agent_name stays 'Hermes Agent'. Adds entry\nto the skins.md docs table and ignores .venv/ in .gitignore.\n","name":"bb/theme","oid":"d30e9b9fb5179cf33749810da417af0bb36232a1"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T17:06:53.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T17:06:53.000000-07:00"},"message":"Port from cline/cline#10343: periodic gateway memory logging\n\nEmit a grep-friendly '[MEMORY] rss=...MB ...' line in agent.log /\ngateway.log every N minutes (default 5) so slow leaks in the long-lived\ngateway process show up as a time series. Based on\nhttps://github.com/cline/cline/pull/10343\n(src/standalone/memory-monitor.ts).\n\n- gateway/memory_monitor.py: new module. Daemon thread, baseline on\n start, final snapshot on stop. Uses resource.getrusage() (stdlib)\n first, falls back to psutil, disables itself with one WARNING if\n neither is available.\n- gateway/run.py: start monitor right after setup_logging() in\n start_gateway(); stop it in the shutdown block next to MCP teardown.\n- hermes_cli/config.py: logging.memory_monitor { enabled, interval_seconds }\n defaults under the existing logging section.\n- tests/gateway/test_memory_monitor.py: 10 unit tests covering format,\n baseline/shutdown snapshots, double-start noop, periodic timer,\n daemon thread invariant, and unavailable-RSS warn-and-skip path.\n\nAdapted from TypeScript/Node to Python (threading.Event-based daemon\nthread instead of setInterval/unref), added Python-specific gc + thread\ncounts to the log line (handier than ext/arrayBuffers for diagnosing\nPython gateway leaks), and gated behind a config.yaml toggle so users\ncan silence the periodic line if they want.\n\nNo heap-snapshot-on-OOM equivalent — CPython doesn't have V8's\n--heapsnapshot-near-heap-limit; tracemalloc would be the Python\nequivalent but adds non-trivial overhead, so leaving that out.\n","name":"cline-port/gateway-memory-monitor","oid":"6366fb9c8be49780a20e1f7751b09c0662d87969"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T17:05:23.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T17:05:23.000000-07:00"},"message":"feat(cli): add /exit --delete flag to remove session on quit\n\nPort from google-gemini/gemini-cli#19332.\n\nUsers can now exit with '/exit --delete' (or '/quit --delete', '/exit -d')\nto permanently remove the current session's SQLite history plus on-disk\ntranscripts (*.json / *.jsonl / request_dump_*) in one shot. Useful for\nprivacy-sensitive workflows and one-off interactions where leaving a\nsession recording behind is undesirable.\n\nImplementation:\n- New HermesCLI._delete_session_on_exit one-shot flag (defaults False).\n- process_command() parses --delete / -d after /exit or /quit and arms\n the flag. Unknown args print a hint and keep the CLI running (prevents\n typos like '/exit -delete' from accidentally exiting).\n- Shutdown path calls SessionDB.delete_session(session_id, sessions_dir=...)\n right after end_session() when the flag is set. That API already\n existed for 'hermes sessions delete' and handles both SQLite removal\n (orphaning child sessions so FK constraints hold) and on-disk file\n cleanup.\n- /quit CommandDef now advertises '[--delete]' in args_hint so /help\n and CLI autocomplete surface it.\n\nTests: tests/cli/test_exit_delete_session.py (12 cases covering both\naliases, case insensitivity, whitespace, short form, unknown-arg\nrejection, and registry metadata).\n\nE2E-verified with isolated HERMES_HOME: session row deleted, all three\ntranscript/request-dump files removed, second delete_session call\ncorrectly returns False.\n","name":"gemini-cli-port/exit-delete-session","oid":"9845650fd69c66cce8a01f6e0b33376f4815b224"},{"author":{"email":"arilotter@gmail.com","name":"ethernet","when":"2026-04-29T19:20:20.000000-04:00"},"committer":{"email":"arilotter@gmail.com","name":"ethernet","when":"2026-04-29T19:20:20.000000-04:00"},"message":"remove relaunch_chat\n\nnot needed\n","name":"fix/relaunch","oid":"c116c74aa43b7eddd2cb6bbb2541a95ff29c6da8"},{"author":{"email":"arilotter@gmail.com","name":"Ari Lotter","when":"2026-04-20T18:52:15.000000-04:00"},"committer":{"email":"arilotter@gmail.com","name":"ethernet","when":"2026-04-29T17:11:30.000000-04:00"},"message":"feat(irc): add interactive setup\n\nfeat(gateway): refine Platform._missing_ and platform-connected dispatch\n\nRestricts plugin-name acceptance to bundled plugin scan + registry\n(no arbitrary string -> enum-pollution), pulls per-platform connectivity\nchecks into a _PLATFORM_CONNECTED_CHECKERS lambda map with a clean\n_is_platform_connected method, and adds tests covering the checker map,\nplugin platform interface, and IRC setup wizard.\n","name":"refactor/gateway-irc","oid":"9cf7c28ba73d2d51c0d7a5d007ed7a07e9cf5c4f"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-29T15:51:42.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-29T15:51:42.000000-05:00"},"message":"fix(tui): swallow empty copy shortcuts\n\nPrevent forwarded copy chords from reaching TextInput as literal input when there is no active selection.\n","name":"bb/tui-copy-on-select","oid":"d3b143c49c95cc4711127c0989e3d7d6e21c679c"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-30T01:07:40.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-30T01:07:53.000000+05:30"},"message":"docs(comfyui): add system detection + hardware-aware recommendations to onboarding\n\nInstead of asking the user what they have, the agent now:\n1. Runs system detection commands (OS, GPU, VRAM, RAM, disk, Python)\n2. Checks if ComfyUI is already installed/running\n3. Recommends the best path based on findings\n\nAdds:\n- Step 1: detection script block (nvidia-smi, system_profiler, etc.)\n- Step 2: decision table mapping detected system → recommended path\n- Hardware requirements table (VRAM tiers, RAM, disk)\n- Specific recommendations per platform:\n macOS → Desktop app, Linux+NVIDIA → comfy-cli, no GPU → Cloud, etc.\n","name":"feat/comfyui-skill-v3","oid":"dac60fb903032c5d238a4254857e24ee4dd4c99e"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T08:18:11.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T08:18:11.000000-07:00"},"message":"fix(cron): action='run' executes inline when no gateway ticker is running\n\nCloses #16612.\n\ncronjob(action='run') and POST /api/jobs/{id}/run previously only called\ntrigger_job(), which just sets next_run_at=now. When no gateway process\nwas running, there was no ticker to pick the job up, so last_run_at\nstayed null forever — the tool returned success=true with nothing\nactually executing.\n\n- cron.scheduler: extract the per-job run+save+deliver+mark pipeline\n from tick()'s closure into a module-level _execute_and_record helper,\n and add run_job_now() which acquires the shared tick file-lock and\n runs one job inline (advancing next_run_at first for recurring jobs\n to preserve at-most-once semantics).\n- tools/cronjob_tools: branch action='run' on gateway presence via\n find_gateway_pids(). Ticker up → defer to next tick (≤60s) with a\n clear message. Ticker down → execute inline via run_job_now() and\n return the updated job including last_run_at/last_status. Tool\n schema description updated to reflect the two modes.\n- gateway/platforms/api_server: same branching on the HTTP endpoint,\n running the inline call off the event loop via run_in_executor.\n- hermes_cli/cron: CLI 'hermes cron run' now surfaces the message,\n last_run_at, and last_status from the tool result instead of\n always printing 'It will run on the next scheduler tick.'\n- docs: update cli-commands and cron-troubleshooting to describe the\n inline-when-no-gateway behaviour.\n\nTests: 11 new unit tests (tests/cron/test_run_job_now.py) covering\nrun_job_now's inline path, _execute_and_record's success/empty/error\npaths, and the tool/API gateway-presence branching. Plus a new\nTestRunJob::test_run_job_executes_inline_when_no_gateway on the\napi_server suite. E2E verified in a temp HERMES_HOME against the\nreal file-based job store.\n","name":"hermes/hermes-dd91e6b2","oid":"9f361cdca6d80c18b1e5ddf07a638116dcb96914"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-28T08:56:12.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-29T20:10:09.000000+05:30"},"message":"feat: provider modules — ProviderProfile ABC, 30 providers, fetch_models, transport single-path\n\nfeat: provider modules — ProviderProfile ABC, 29 providers, fetch_models, transport single-path\n\nIntroduces providers/ as the single source of truth for every inference\nprovider. All 29 providers declared with correct data cross-checked against\nauth.py, runtime_provider.py and auxiliary_client.py.\n\nRebased onto main (30307a980). Incorporates post-salvage fixes from\n56724147e (gmi aux model google/gemini-3.1-flash-lite-preview, already set in providers/gmi.py).\n","name":"feat/provider-modules","oid":"84d1673e2fa829755fcffa0b09c6cb323c11a2f7"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T04:41:43.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-29T04:41:43.000000-07:00"},"message":"perf(tools): cache built-in tool module list in tools/_manifest.py\n\nSkips the ~145 ms AST scan of every tools/*.py file at every startup.\nCLI cold import drops ~100 ms on a machine where Phase 1's other\nlazy-imports already landed.\n\n## What changed\n\ntools/registry.py:_load_manifest() reads a committed tuple of tool\nmodule names from tools/_manifest.py. discover_builtin_tools() tries\nthe manifest first; falls back to the AST scan only when the manifest\nis missing (fresh checkout, or the generator hasn't been run yet) or\nwhen the caller passes a non-default tools_dir (tests, embedders).\n\nscripts/build_tool_manifest.py regenerates the manifest. Run manually\nafter adding a new tools/*.py, or let CI catch drift.\n\n.github/workflows/tests.yml runs the generator with --check on every\nPR. Rejects commits that add/remove a self-registering tool module\nwithout updating the manifest.\n\ntools/_manifest.py is the committed artifact, 27 built-in modules.\n\n## Design decisions\n\nNo runtime mtime-drift check. Considered adding a stat walk of\ntools/*.py at startup to auto-invalidate the manifest on dev machines,\nrejected because:\n - Adds overhead to the path we're trying to speed up (66 stats,\n ~5-10 ms worst case, defeats part of the savings).\n - Gives false positives when devs edit tools/ helpers (ansi_strip,\n approval, budget_config, path_security, etc.) that don't call\n registry.register() — those edits wouldn't require a regen.\n - CI check is strictly better: catches the real error class (manifest\n missing a self-registering tool) without runtime cost, and developers\n get feedback at PR time instead of silent divergence.\n\nManifest coverage: built-in tools/*.py ONLY. Plugin tools (via\nctx.register_tool) and MCP tools (registered dynamically on server\nconnect) use separate registration paths that this change doesn't\ntouch. End-user-visible behavior: unchanged — installing a plugin or\nconnecting an MCP server works exactly as before.\n\n## Failure modes\n\n1. Manifest lists a tool that no longer exists → importlib.import_module\n raises, logged as warning, skipped. Non-fatal but noisy. CI check\n catches this class before merge.\n\n2. Manifest missing a new tool → the tool isn't registered at startup.\n User-visible. CI --check catches before merge.\n\n3. Manifest missing entirely (tools/_manifest.py deleted) → fallback to\n AST scan, startup gets slower by ~145 ms but behavior is identical.\n\n## Measurements\n\ndiscover_builtin_tools() isolated:\n without manifest (AST scan): 674 ms min, 702 ms median\n with manifest: 524 ms min, 550 ms median\n savings: 150 ms\n\nFull 'import cli':\n before (main): ~950 ms\n after (this PR): ~846 ms\n savings: ~100 ms\n\n(The isolated savings is bigger than the cli delta because some AST\nscan overhead was being double-paid via transitive imports; manifest\nremoves it once at the source.)\n\n## Validation\n\n- 121/121 tests/tools/test_registry + test_terminal_tool +\n test_skills_tool pass (unchanged from main).\n- 68 tools registered with manifest (matches baseline).\n- CI --check flag verified: passes for the committed manifest, fails\n with exit 1 when I drop a new tool into tools/ without regenerating.\n- Live hermes chat smoke: 1 turn + /quit, agent successfully enumerated\n all registered tools when asked to list them. 36-line log window,\n zero errors.\n","name":"hermes/hermes-dab6fbf1","oid":"53f875d063ad8d13d25589777db7e6478ded6e84"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-29T13:50:17.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-29T13:50:17.000000+05:30"},"message":"style(comfyui): format SKILL.md — table alignment, YAML tags, onboarding hint\n","name":"feat/comfyui-skill-v2","oid":"4d810d959157e350a85011d9f2e137e3d4696055"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"brooklyn!","when":"2026-04-28T22:45:14.000000-07:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-04-28T22:45:14.000000-07:00"},"message":"Merge pull request #17190 from NousResearch/bb/tui-cold-start-profiling\n\nperf(tui): cut visible cold start ~57% with lazy agent init","name":"vbrunet/2026_04_29-terminal-titles","oid":"5e68503d2f90aa2822494c910809eae1c3895238"},{"author":{"email":"teknium@users.noreply.github.com","name":"teknium1","when":"2026-04-28T18:34:30.000000-07:00"},"committer":{"email":"teknium@users.noreply.github.com","name":"teknium1","when":"2026-04-28T18:34:30.000000-07:00"},"message":"feat(kanban): dispatcher runs in the gateway by default; retire standalone daemon\n\nThe kanban dispatcher is now embedded in the gateway process —\n'gateway is the single dispatcher host' — removing the need for a\nseparate `hermes kanban daemon` or systemd unit. Typical user path\nbecomes: `hermes gateway start` + `hermes kanban create ...` and\nready tasks run on the next tick (60s default). The cost is ~300µs per\ninterval on an idle board; negligible.\n\nChanges:\n- config: new `kanban` section in DEFAULT_CONFIG with\n `dispatch_in_gateway: true` (default on) and\n `dispatch_interval_seconds: 60`. Additive — no \\_config_version bump.\n- gateway/run.py: new `_kanban_dispatcher_watcher()` background task,\n symmetric with `_kanban_notifier_watcher`. Reads config at boot;\n exits cleanly if the flag is off. Runs each tick via\n `asyncio.to_thread` so the SQLite WAL lock never blocks the loop.\n Sleeps in 1s slices so shutdown is snappy. Health telemetry mirrored\n from `_cmd_daemon` — warns when the ready queue is non-empty for\n 6 consecutive ticks with 0 spawns. Env override\n `HERMES_KANBAN_DISPATCH_IN_GATEWAY=0` disables without editing\n config.yaml.\n- hermes_cli/kanban.py: `_cmd_daemon` becomes a deprecation stub — no\n `--force`, exits 2 with migration guidance pointing at\n `hermes gateway start`. With `--force` (undocumented, help=SUPPRESS)\n the old standalone loop still runs, for headless hosts that can't run\n the gateway. Help text marks the subcommand DEPRECATED.\n- hermes_cli/kanban.py: new `_check_dispatcher_presence()` helper —\n returns (running, human_message) by probing `gateway.status.get_running_pid`\n AND reading `kanban.dispatch_in_gateway`. Defensive: import/probe\n failures return (True, \"\") so we never cry wolf on partial installs.\n- `hermes kanban create` prints a stderr warning when the task lands\n in 'ready' with an assignee but no dispatcher will pick it up. Skipped\n in `--json` mode so machine-parseable output stays clean. Skipped for\n triage/unassigned tasks (they can't dispatch regardless).\n- `hermes kanban init` prints `hermes gateway start` as the next step\n (was: `hermes kanban daemon`).\n- plugin_api.py: POST /tasks response includes a `warning` field when\n the same probe returns not-running, so the dashboard UI can surface\n a banner. Probe errors are swallowed silently — must never break\n create.\n- dashboard dist/index.js: `createTask` threads the `warning` response\n field into the existing error-banner channel with 'Task created,\n but: ...'.\n- toolsets.py: \"spawned by `hermes kanban daemon`\" description updated\n to \"spawned by the kanban dispatcher (gateway-embedded by default)\".\n Matching change to the tools/kanban_tools.py docstring.\n- systemd unit: marked DEPRECATED in its Description + header comment,\n invokes the standalone daemon via the explicit `--force` flag so\n users who haven't migrated don't accidentally spawn duplicate\n dispatchers.\n- docs: kanban.md 'Running the dispatcher as a service' section\n rewritten to describe gateway-embedded dispatch as the default; the\n standalone systemd instructions are gone. kanban-tutorial.md 'Start\n the daemon' block replaced with `hermes gateway start`. CLI command\n reference table now marks `hermes kanban daemon` DEPRECATED.\n\nTests (17 new):\n- DEFAULT_CONFIG has kanban.dispatch_in_gateway=True and a sane\n interval.\n- `_check_dispatcher_presence` returns running when gateway pid is\n found and flag is on; warns with `hermes gateway start` guidance\n when no gateway; warns with `dispatch_in_gateway` guidance when\n flag is off; silent on probe error.\n- `hermes kanban create` (non-JSON) warns on stderr when no gateway\n and task is ready+assigned; stays silent when gateway is up; never\n warns on triage or unassigned tasks.\n- `hermes kanban daemon` without --force prints DEPRECATED + exits 2.\n- Argparse help for `kanban daemon` contains the word DEPRECATED.\n- Gateway `_kanban_dispatcher_watcher` respects config flag=False\n (exits fast), respects HERMES_KANBAN_DISPATCH_IN_GATEWAY=0 env\n override, and treats truthy env as 'defer to config' (not force-on).\n- Dashboard POST /tasks response carries `warning` when probe says\n no dispatcher; omits it when probe says running; skips probe on\n triage tasks; survives probe errors without breaking create.\n\nE2E verified in an isolated HERMES_HOME: watcher exits in 0.000s when\nflag is off, 0.000s on env override, gracefully stops within 3s of\n`_running=False` with the flag on.\n\nIntegration issues fixed in the same pass:\n- Stale 'hermes kanban daemon --assignee' refs in kanban-tutorial.md\n (that flag never existed; just bad author copy).\n- Stale toolset description pointing at the retired daemon.\n- Stale docstring in tools/kanban_tools.py for `_check_kanban_mode`.\n- Systemd unit still invoking `hermes kanban daemon` without\n `--force`; now invokes with `--force` and is marked DEPRECATED.\n","name":"feat/kanban-standing","oid":"60e75674e7f19c0a50cd3edfb256fc72c8ac1072"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-28T15:52:41.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-28T17:54:39.000000-05:00"},"message":"review(copilot): point reload.env docstring at hermes_cli.config.reload_env\n","name":"bb/tui-reload-env","oid":"059652d11aba899665ad0720ce2a20d7d68c9726"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-27T15:08:44.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-28T17:01:34.000000-05:00"},"message":"fix(tui): improve learning ledger scanability\n\nShrink the details panel share and simplify category row labels so the four learning lists are easier to scan.\n","name":"bb/learning-ledger","oid":"f45b844670e6a329698b6064e8134ba35e3e1c99"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-28T15:30:11.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-28T15:30:11.000000-05:00"},"message":"review(copilot): keep separator space when verb hits pad limit + reword comment\n","name":"bb/tui-status-ticker-width","oid":"4d0acda13475574909cd66fa43540f8e51725b49"},{"author":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-28T13:24:29.000000-04:00"},"committer":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-28T13:24:29.000000-04:00"},"message":"fix(plugins): register dynamically-loaded modules in sys.modules before exec\n\nDashboard plugin API routes (web_server._mount_plugin_api_routes) and\ngateway event hooks (gateway.hooks.HookRegistry.discover_and_load) both\nloaded Python files via importlib.util.spec_from_file_location +\nexec_module without registering the resulting module in sys.modules.\n\nThat breaks any plugin or hook handler that uses `from __future__ import\nannotations` together with a Pydantic BaseModel / dataclass / anything\nthat introspects `__module__`: at first request Pydantic tries to\nresolve string-form type hints against the defining module's namespace,\ncan't find it by name, and raises:\n\n PydanticUserError: TypeAdapter[...] is not fully defined;\n you should define ... and all referenced types,\n then call `.rebuild()` on the instance.\n\nThis is what broke the kanban dashboard's 'triage' button — POST\n/api/plugins/kanban/tasks validated against CreateTaskBody (a Pydantic\nmodel in a file using `from __future__ import annotations`) and\nreturned 500 on every click.\n\nThe fix, applied symmetrically to both loaders:\n\n 1. Compute module_name once.\n 2. Register the module in sys.modules BEFORE exec_module.\n 3. On exec_module failure, pop the half-initialized stub so subsequent\n reloads don't pick up broken state.\n\nGETs were unaffected because they don't build a body TypeAdapter, which\nis why this only surfaced when users started POSTing.\n","name":"fix/plugin-loader-sys-modules-registration","oid":"854206e59e1a3fa6e5184c982746401a76157656"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-28T19:21:35.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-28T19:33:26.000000+05:30"},"message":"fix(nix): wire extraPackages into per-user profile for terminal visibility\n\nextraPackages previously only added packages to the systemd service\nPATH. The terminal backend's login-shell snapshot rebuilds PATH from\nNixOS system profiles (/etc/set-environment), so tools added via\nextraPackages were invisible to terminal commands, skills, and cron\njobs — the entire use case.\n\nFix: when extraPackages is non-empty, also set\nusers.users.${cfg.user}.packages, which NixOS wires into\n/etc/profiles/per-user//bin — a path that IS included in the\nlogin-shell PATH. The systemd service PATH still includes them too,\nso both the process and spawned commands see the packages.\n\nAlso updates docs to reflect the corrected behavior.\n","name":"fix/deprecate-extrapackages-nix","oid":"4e4f55a83e24e868803539fa31c42ab7d472c190"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-28T01:15:46.000000-07:00"},"committer":{"email":"teknium@users.noreply.github.com","name":"teknium1","when":"2026-04-28T02:13:06.000000-07:00"},"message":"fix(computer-use): harden image-rejection fallback + AUTHOR_MAP\n\nFollow-up to #15328's vision-unsupported retry branch in run_agent.py.\n\n_strip_images_from_messages() previously deleted any message whose content\nwas entirely images. That's fine for synthetic user messages injected for\nattachment delivery, but it breaks providers for tool-role messages — the\npaired tool_call_id on the preceding assistant message ends up unmatched,\nwhich OpenAI-compatible APIs reject with HTTP 400.\n\nFix: tool-role messages whose content becomes empty are replaced with a\nplaintext placeholder that preserves the tool_call_id linkage. Only\nnon-tool messages are dropped. Added 10 tests covering the role-alternation\ninvariants + image-type coverage.\n\nImage-rejection detector: expanded phrase list (image content not\nsupported / multimodal input / vision input / model does not support\nimage) and gated on 4xx status so transient 5xx errors never get\nmisinterpreted as 'server said no to images'. Detection is documented as\nbest-effort English phrase matching.\n\nAUTHOR_MAP: mapped 3820588+ddupont808@users.noreply.github.com to\nddupont808 so release notes attribute the salvage correctly.\n","name":"hermes/hermes-c8604b32","oid":"91391a6811543731c82b9572e45aaf676aaf4bda"},{"author":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-27T22:29:17.000000-04:00"},"committer":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-28T02:40:25.000000-04:00"},"message":"fix(dashboard): persist chat tab state across tab switches\n\nThe dashboard's Chat tab (hermes dashboard --tui) lost its session\nwhenever the user navigated to another tab and came back. React Router\nunmounted ChatPage on path change, which ran the cleanup function,\nclosed the PTY WebSocket, and terminated the underlying TUI child -\nso the next mount generated a fresh channel id, spawned a new PTY, and\nstarted a brand-new conversation.\n\nRather than rebuild the destroyed state (session id capture + resume\nvia HERMES_TUI_RESUME would reload history from disk but drop in-flight\ntool state, scrollback, and picker position), keep the component tree\nalive.\n\n* Pull ChatPage out of Routes into a sibling always-mounted host that\n toggles visibility via display:none keyed off the current route. A\n tiny ChatRouteSink still claims /chat so the catch-all redirect\n does not fire.\n* xterm instance, WebSocket, PTY child, and TUI/agent state all\n survive; returning to /chat shows the exact conversation the user\n left.\n* Respect plugin `/chat` overrides: if a plugin manifest declares\n `tab.override: \"/chat\"`, the Routes tree already swaps the element\n for — we additionally suppress the persistent host\n so the two don't paint on top of each other. Preserves the\n pre-persistence contract that a plugin owning /chat replaces the\n built-in chat UI entirely.\n* Wait for usePlugins() to finish loading before mounting the\n persistent host. Manifests arrive asynchronously from\n /api/dashboard/plugins, so without the `!pluginsLoading` gate the\n host would mount with manifests=[], spawn a PTY, and then unmount\n mid-session when the manifest list resolves and reveals a /chat\n override. Typical delay is <50ms; worst case is the 2s plugin-\n registration safety timeout. Cheaper than killing someone's\n conversation underneath them.\n* Gate page-header slot (`setEnd`), the mobile sheet's portalled\n render, and body-scroll lock on a new `isActive` prop so the hidden\n ChatPage doesn't fight the active page for shared state. The\n scroll-lock effect keys on the *derived* `mobilePanelOpen` (which is\n `isActive && mobilePanelOpenRaw`) rather than the raw state — that\n way tab-switch flips the dep false, fires the cleanup, and releases\n `document.body.style.overflow`. Keying on the raw state would leave\n body.overflow=\"hidden\" stuck on /sessions and every other tab until\n the user navigated back to /chat and explicitly closed the sheet.\n* When isActive flips false to true, force a double-rAF fit:\n display:none collapses the host box and ResizeObserver does not fire\n on display changes, so xterm would otherwise stay at a stale or 1x1\n grid. Also early-return from syncTerminalMetrics when the host has\n zero area, since fit() on a zero-sized element produces a 1x1\n terminal.\n* Focus handling on tab return: only steal focus into the terminal if\n focus wasn't already parked somewhere inside ChatPage (e.g. the\n sidebar model picker, a tool-call entry). Yanking focus away from\n whatever the user last clicked is surprising and a screen-reader\n foot-gun; the typical \"first activation\" case still focuses the\n terminal because document.activeElement is at that point.\n\nTrade-off worth flagging, deliberately not mitigated in this change:\nwhile hidden, ChatPage still holds a PTY child + WebSocket + xterm\ninstance for the dashboard's full lifetime. The WS keeps delivering\nbytes and xterm keeps parsing them into a display:none host (cheap —\nno paint work, but not free). Reasonable costs to pay for the session\npreservation; if they become a problem we can pause `term.write` when\n!isActive or idle-disconnect after N minutes hidden.\n\nLint clean on touched files. tsc -b && vite build pass.\n","name":"fix/chat-tab-persistence","oid":"ddd2542ba590dd16e72dbf7c12741d6984a89d4b"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-28T11:36:46.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-28T11:52:42.000000+05:30"},"message":"feat(plugins): add optional-plugins/ discovery + langfuse_tracing as first official optional plugin\n\nIntroduces optional-plugins/ — a new category for plugins that ship with\nthe repo but are NOT auto-discovered. They live alongside the code but only\nland in ~/.hermes/plugins/ (and thus get loaded) when the user explicitly\ninstalls them.\n\nCore changes:\n- optional-plugins/observability/langfuse-tracing/ — langfuse tracing plugin\n (pre/post LLM + tool hooks, usage/cost normalization, fail-open when SDK\n missing). NOT in plugins/ so zero import overhead on devices that don't\n want it.\n- hermes_cli/plugins_cmd.py — official install path: _resolve_official_plugin()\n recognises 'official//' identifiers and copies from\n optional-plugins/ into ~/.hermes/plugins/ (no git clone, no network).\n _list_official_plugins() enumerates available optional plugins.\n cmd_list(available=True) shows not-yet-installed official plugins.\n- hermes_cli/main.py — hermes plugins list --available flag\n- hermes_cli/tools_config.py — Langfuse Observability in TOOL_CATEGORIES;\n post_setup handler installs the langfuse SDK and runs cmd_install()\n- hermes_cli/config.py — Langfuse credentials in OPTIONAL_ENV_VARS;\n optional tuning keys in _EXTRA_ENV_KEYS\n\nUser flows:\n hermes plugins install official/observability/langfuse-tracing\n hermes plugins list --available\n hermes tools (-> Langfuse Observability -> credentials -> auto-installs)\n\nCloses #15764\n","name":"feat/langfuse-plugin","oid":"d8c2c77be6f44baf675c5ecd2533c81482348a1e"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-27T17:05:44.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-27T17:05:44.000000-07:00"},"message":"Port from Kilo-Org/kilocode#9448: roll up subagent costs into parent session total\n\nChild subagents built by delegate_task() each track their own\nsession_estimated_cost_usd, but the parent agent's total never folded\nthose numbers in. On runs where the parent mostly delegates and the\nchildren do the expensive work, the footer/UI was reporting a fraction\nof the actual spend — sometimes $0.00 when the parent itself made no\nbilled calls.\n\nFix:\n- Capture each child's session_estimated_cost_usd into _child_cost_usd\n on the result entry (before child.close() drops the counter).\n- After the existing subagent_stop hook loop, sum the children's costs\n and add the total to parent.session_estimated_cost_usd.\n- Promote session_cost_source from 'none' -> 'subagent' when the parent\n had no direct spend but children did, so the UI doesn't label the\n total as having unknown provenance. Real sources (openrouter,\n anthropic, etc.) are preserved.\n\nNested orchestrator -> worker trees roll up naturally: each layer's own\ndelegate_task() folds its direct children in, and when the orchestrator\nitself returns, its parent folds the orchestrator's now-inflated total\non top.\n\nInternal fields (_child_cost_usd, _child_role) are stripped from the\nresults dict before it's serialised back to the model — same contract\nas _child_role already followed.\n\nTests: TestSubagentCostRollup (5 cases) covers single-child, batch,\nzero-cost-children, preserved-source, and legacy-fixture paths.\n\nSource: https://github.com/Kilo-Org/kilocode/pull/9448\n","name":"kilocode-port/subagent-cost-rollup","oid":"dad42014aeb11bf7e655868f00d1c295875dddc1"},{"author":{"email":"jonny@nousresearch.com","name":"yoniebans","when":"2026-04-27T19:14:30.000000+02:00"},"committer":{"email":"morpheus_hermes@proton.me","name":"morpheus","when":"2026-04-27T21:48:41.000000+02:00"},"message":"fix: include cache tokens in dashboard analytics input totals\n\nThe /api/analytics/usage endpoint summed only the raw input_tokens\ncolumn, which for Anthropic-direct sessions holds only the uncached\nportion of the prompt. cache_read_tokens and cache_write_tokens\n(which complete the total prompt) were ignored.\n\nThis caused the dashboard to massively undercount token usage —\nshowing ~117M instead of ~345M over 30 days — since Anthropic\nsessions with high cache hit rates stored almost all prompt tokens\nin the cache columns.\n\nFix: fold COALESCE(cache_read_tokens, 0) + COALESCE(cache_write_tokens, 0)\ninto the input_tokens sum across all three SQL queries (daily, by-model,\ntotals). This is correct for every provider because normalize_usage()\nguarantees input_tokens + cache_read + cache_write = total prompt tokens\nregardless of API shape (Anthropic / OpenAI / Codex).\n\nAdd a regression test that creates a session with Anthropic-style token\nsplits and asserts the endpoint returns the combined total.\n","name":"fix/analytics-include-cache-tokens","oid":"18e6cd993861ac8e1c109a36d1e9f60d403987e0"},{"author":{"email":"51058514+Sanjays2402@users.noreply.github.com","name":"Sanjay Santhanam","when":"2026-04-26T22:08:25.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T22:08:25.000000-07:00"},"message":"fix(terminal): suppress declare -x stdout leak when sourcing snapshot on macOS\n\nOn macOS (bash 3.2 and certain Homebrew builds), sourcing a file\ncontaining declare -x statements prints each declaration to stdout.\nThe session snapshot (written by export -p) is full of these, so\nevery command execution prepends ~60 lines of env vars into the LLM's\ncontext window, wasting tokens and potentially exceeding context\nlimits (issue #15459).\n\nFix: redirect both stdout and stderr to /dev/null when sourcing the\nsnapshot. The source builtin still sets the variables in the current\nshell — only the print output is suppressed. On Linux this is already\nsilent, so the redirect is harmless.\n\nFixes #15459\n","name":"hermes/hermes-8fedd55b","oid":"a52ed706085f2e8f2c0d19d69aa727224eade716"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T23:31:49.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T23:31:49.000000-05:00"},"message":"Fix TUI input field ANSI leaks and text selection issues\n\n1. Fix ANSI escape code leakage during scroll operations:\n - Add ensureSafeAnsi utility to ensure all ANSI sequences are properly terminated\n - Modify renderWithCursor and renderWithSelection to always include reset codes\n - Add final safety check in text rendering to catch any potential leaks\n\n2. Improve text selection in input field:\n - Add proper mouse drag event handling for text selection\n - Enhance click handlers to support selection operations\n - Fix edge cases in selection rendering\n - Ensure proper reset of ANSI codes after selections\n\nFixes reported issues where users couldn't copy/paste text in input field and\nexperienced ANSI leakage during scrolling operations.\n","name":"bb/fix-tui-input-selection","oid":"b9393296fcc801869f6957a45aa88c5bcc23e12c"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T23:19:50.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T23:19:50.000000-05:00"},"message":"Merge origin/main into bb/p2-mru-resume-order\n","name":"bb/p2-mru-resume-order","oid":"b4bc87bfab57473808cac3a3fbdd49213cafc701"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T21:18:59.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T21:18:59.000000-07:00"},"message":"fix(feishu): send WebSocket CLOSE frame on disconnect (#10202)\n\nFeishu adapter's disconnect() cancelled WSS-thread tasks but never\ncalled the lark_oapi client's _disconnect() coroutine, so no\nWebSocket CLOSE frame was sent. Feishu's server kept routing\nmessages to the stale endpoint for minutes (CLOSE-WAIT timeout),\nsilencing the channel across every shutdown path — systemd restart,\nhermes update, hermes gateway restart, and the --replace takeover\nduring 'hermes dashboard' invocations.\n\nSchedule ws_client._disconnect() on the WSS thread loop via\nrun_coroutine_threadsafe with a 5s timeout before the existing\ntask-cancel + loop-stop sequence. Defensive hasattr guard + broad\nexcept keeps disconnect() resilient if lark_oapi's internals shift.\n\nFixes #10202\n","name":"hermes/hermes-2c880154","oid":"cbf66fcfcd4f44b77eab0a52189bbe528cc28daa"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T19:52:17.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T19:52:17.000000-07:00"},"message":"fix(tests): update copilot_acp subprocess.Popen patch paths to acp_adapter.copilot_client\n\nMain added two HOME-handling tests (test_run_prompt_prefers_profile_home_when_available,\ntest_run_prompt_passes_home_when_parent_env_is_clean) after PR #14424 was written.\nThese patch 'agent.copilot_acp_client.subprocess.Popen', but the shim module no longer\nhas 'subprocess' imported. Update patch strings to target the real module location.\n\nFollow-up commit on the salvage PR; kshitijk4poor's original commit is preserved above.\n","name":"hermes/hermes-8b414b52","oid":"5dbd97bdfb165c3bc9a8cbdd74dcc50f7cbf1c94"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T17:04:57.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T17:04:57.000000-07:00"},"message":"fix(signal): read groupV2.id in envelope, fall back to legacy groupInfo\n\nPort from qwibitai/nanoclaw#1962: modern Signal V2-only groups surface on\ndataMessage.groupV2.id, not groupInfo.groupId. signal-cli versions differ\nin which field they expose for V2 groups — some forward the underlying\nlibsignal envelope verbatim (groupV2), others normalize everything into\ngroupInfo. Without a groupV2 read, V2-only groups appear as DMs because\ngroupInfo is undefined and the adapter misroutes them to the sender's\nDM session.\n\nReads groupV2.id first, falls back to groupInfo.groupId. Also hardens\nchat_name extraction against non-dict groupInfo payloads (crashed with\nAttributeError under malformed envelopes).\n\n6 new tests cover V2 routing, V1 legacy compatibility, V2-preferred\nprecedence, no-group DM path, allowlist enforcement, and malformed\npayloads.\n","name":"nanoclaw-port/signal-groupv2","oid":"61b65eeb0c93710e07e3e510a935196c93cff72f"},{"author":{"email":"4317663+helix4u@users.noreply.github.com","name":"helix4u","when":"2026-04-25T16:43:55.000000-06:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T12:55:06.000000-07:00"},"message":"fix(tui): bound live render memory pressure\n","name":"hermes/hermes-d7874f79","oid":"1d24cb0e6e430c6b506c27e0a9dda929bb9e4477"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T07:16:35.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-26T07:16:35.000000-07:00"},"message":"Merge remote-tracking branch 'origin/main' into hermes/curator-infra\n","name":"hermes/curator-infra","oid":"13ca9ee665bb3dda2b728d7082d8e68d5ab51b1d"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T01:27:55.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T01:27:55.000000-05:00"},"message":"feat(tui): implement optimized transcript pane with virtualization\n\nReplace the standard ScrollBox with a new OptimizedTranscriptPane component\nthat uses the FixedWindowScroller for virtualized message rendering:\n\n1. Implement OptimizedTranscriptPane as a drop-in replacement\n - Preserves all existing functionality\n - Maintains same rendering logic for messages\n - Uses efficient virtualization under the hood\n\n2. Integrate OptimizedTranscriptPane with appLayout\n - Enable performance mode by default\n - Preserve layout and scrollbar positioning\n - Keep the original implementation as fallback\n\nThis completes the TUI performance optimizations for long sessions,\naddressing scrolling lag and input jitter by dramatically reducing\nDOM nodes and layout calculations.\n","name":"bb/tui-perf-fix","oid":"75b7bad6be5661dd4331bb8884d4586b3fd21ce8"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T01:17:21.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-26T01:17:21.000000-05:00"},"message":"feat(tui): add model picker and approval/clarify prompt workflows\n\nThree new static prompt components for showroom capture (no useInput):\n\n- ApprovalPromptStatic: double-bordered warning box with command preview,\n 4 options (allow once/session/always/deny), ▸ selection, footer hints\n- ClarifyPromptStatic: heading + numbered choices with 'Other' option\n- ModelPickerStatic: double-bordered popup with provider/model lists,\n current model header, persist toggle, quick-pick numbers\n\nNew workflows:\n- interactive-prompts: approval → clarify → deploy result\n- model-picker: provider stage → model stage → switch result\n","name":"feature/tui-showroom","oid":"036dd14425d5012e6beefc78cd80942dd68b858e"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-25T19:48:02.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-25T19:48:02.000000-05:00"},"message":"feat(gui): make desktop setup flow real and testable\n\nAdd a GUI-first setup gate and runtime state API so desktop onboarding is safe, iterative, and works with isolated fresh-mode installs. Scaffold and wire the desktop shell/runtime pieces so this branch runs end-to-end without disturbing existing user installs.\n","name":"bb/base-gui","oid":"648da6a8d1c13fd259ea46575c47b8f4e2210a17"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-25T12:41:57.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-25T12:41:57.000000-05:00"},"message":"fix(tui): rebuild when Ink bundle is missing\n","name":"investigate/fix-tui-container-ink-bundle","oid":"d603bce4cbe139e1d62a53b3dc155b1e5649e062"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T19:31:27.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T19:53:54.000000+05:30"},"message":"fix(flush_memories): always deduct headroom + resolve flush aux model + trim defence\n\nThree fixes for flush_memories / compression context window overflow:\n\n1. ALWAYS deduct headroom before comparing aux_context vs threshold.\n #15631 only deducted inside 'if aux_context < threshold' — which\n never fires in the common same-model case (threshold = context × 0.50\n means aux_context > threshold always). Now headroom is computed\n unconditionally and effective_limit = aux_context - headroom is\n compared against threshold.\n\n2. Also resolve flush_memories auxiliary model in the feasibility check.\n If the user configures separate auxiliary.flush_memories provider,\n the flush model's smaller context was unchecked.\n\n3. Defence-in-depth trimming in flush_memories() for CLI /new and\n gateway resets that bypass preflight compression entirely.\n","name":"fix/flush-memories-context-overflow","oid":"25d9fc8094c2cf398b6dc3afc2ba1dc7d286aeb2"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-24T07:21:09.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-25T06:17:44.000000-07:00"},"message":"feat: compression eval harness for agent/context_compressor.py\n\nShips a complete offline eval harness at scripts/compression_eval/. Runs\na real conversation fixture through ContextCompressor.compress(), asks\nthe compressor model to answer probe questions from the compressed\nstate, then has a judge model score each answer 0-5 on six dimensions\n(accuracy, context_awareness, artifact_trail, completeness, continuity,\ninstruction_following). Methodology adapted from Factory's Dec 2025\nwrite-up (https://factory.ai/news/evaluating-compression); the\nscoreboard framing is not adopted.\n\nMotivation: we edit context_compressor.py prompts and _template_sections\nby hand and ship with no automated check that compression still\npreserves file paths, error codes, or the active task. Until now there\nhas been no signal between 'test suite green' and 'a user hits a bad\nsummary in production.'\n\nWhat's shipped\n- DESIGN.md — full architecture, fixture/probe format, scrubber\n pipeline, grading rubric, open follow-ups\n- README.md — usage, cost expectations, when to run it\n- scrub_fixtures.py — reproducible pipeline that converts real sessions\n from ~/.hermes/sessions/*.jsonl into public-safe JSON fixtures. Applies\n agent.redact.redact_sensitive_text + username path normalisation +\n personal handle scrubbing + email/git-author normalisation + reasoning\n scratchpad stripping + platform-mention scrubbing + first-user\n paraphrase + system-prompt placeholder + orphan-message pruning + 2KB\n tool-output truncation\n- fixtures/ — three scrubbed session snapshots covering three session\n shapes:\n feature-impl-context-priority (75 msgs / ~17k tokens)\n debug-session-feishu-id-model (59 msgs / ~13k tokens)\n config-build-competitive-scouts (61 msgs / ~23k tokens)\n- probes/ — three probe banks (10-11 probes each) covering all four\n types (recall/artifact/continuation/decision) with expected_facts\n anchors (PR numbers, file paths, error codes, commands)\n- rubric.py — six-dimension grading rubric, judge-prompt builder,\n JSON-with-fallback response parser\n- compressor_driver.py — thin wrapper around ContextCompressor for\n forced single-shot compression (fixtures are below the default\n 100k threshold so we force compress() to attribute score deltas\n to prompt changes, not threshold-fire variance)\n- grader.py — two-phase continuation + grading calls via the OpenAI\n SDK directly against the resolved provider endpoint\n- report.py — markdown report renderer (paste-ready for PR bodies),\n --compare-to delta mode, per-run JSON dumper\n- run_eval.py — fire-style CLI (--fixtures, --runs, --judge-model,\n --compressor-model, --label, --focus-topic, --compare-to, --verbose)\n- tests/scripts/test_compression_eval.py — 33 hermetic unit tests\n covering rubric parsing edge cases, judge-prompt building, report\n rendering, summariser medians, per-run JSON roundtrip, fixture and\n probe loading, and a PII smoke check on the checked-in fixtures\n\nNon-LLM paths are covered by the 33-test suite that runs in CI. The\nLLM paths (continuation + grading) require credentials and real API\ncalls, so they're exercised by running the eval itself — not by CI.\n\nValidation\n- 33/33 unit tests pass in 0.33s via scripts/run_tests.sh\n- 50/50 adjacent tests (tests/agent/test_context_compressor.py) still\n pass — no regression introduced\n- End-to-end dry run against debug-session-feishu-id-model with\n openai/gpt-5.4-mini via Nous Portal:\n Compression: 13081 -> 3055 tokens (76.6% ratio), 59 -> 10 messages\n Overall score: 3.25 (artifact_trail 1.50 is the weak spot,\n matching Factory's published observation)\n Specific probe misses surfaced with concrete judge notes\n\nNoise floor (one empirical data point)\nSame inputs re-run: overall 3.25 -> 3.17 (delta -0.08). Individual\ndimensions varied up to ±0.5 between two single-run medians. Confirms\nthe DESIGN.md < 0.3 noise guidance is the right order of magnitude\nfor single-run comparisons. Tighter noise measurement (N=10) is\ntracked as an open follow-up in DESIGN.md.\n\nWhy scripts/ and not tests/\nRequires API credentials, costs ~$0.50-1.50 per run, minutes to\nexecute, LLM-graded (non-deterministic). Incompatible with\nscripts/run_tests.sh which is hermetic, parallel, credential-free.\nscripts/sample_and_compress.py is the existing precedent for offline\ncredentialed tooling.\n\nOpen follow-ups (tracked in DESIGN.md, not blocking this PR)\n1. Iterative-merge fixture (two chained compressions on one session)\n2. Precise noise-floor measurement at N=10\n3. Scripted scrubber helpers to lower the cost of fixture #4+\n4. Judge model selection policy (pin vs. per-user)\n","name":"design/compression-eval-harness","oid":"1e6285c53db6f96fbb57f1ecb38def158a04fb01"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T17:44:32.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T17:58:24.000000+05:30"},"message":"fix: /stop now immediately aborts streaming retry loop\n\nWhen a user sends /stop during a streaming API call, the outer poll loop\ndetects _interrupt_requested and closes the HTTP connection. However, the\ninner _call() thread catches the connection error and enters its retry\nloop — opening a FRESH connection without checking the interrupt flag.\n\nOn slow providers like ollama-cloud, each retry attempt blocks for the\nfull stream-read timeout (120s+). With 3 retry attempts this caused\n510+ second delays between /stop and actual response — the agent appeared\ncompletely unresponsive despite the stop being acknowledged.\n\nFix: add an _interrupt_requested check at the top of the streaming retry\nloop so the agent exits immediately instead of retrying.\n\nAlso fix log truncation: all session key logging in gateway/run.py used\n[:20] or [:30] slices, which truncated 'agent:main:telegram:dm:5690190437'\n(33 chars) to 'agent:main:telegram:' — losing the identifying chat type\nand user ID. Replace with full keys to make logs debuggable.\n\nReported by user Sidharth Pulipaka via Telegram on ollama-cloud provider.\n","name":"fix/stop-interrupt-retry-loop","oid":"2579245ae25fc36d966f026bcef6ba194957e022"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T14:10:42.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-25T14:10:42.000000+05:30"},"message":"fix(model): preserve custom endpoint credentials and accept cloud models not in /v1/models\n\nWhen switching models on a custom endpoint (ollama-launch):\n- Same-provider switches no longer re-resolve credentials (fixes base_url\n being lost for 'custom' provider on subsequent switches)\n- Named providers (ollama-launch) are resolved via user_providers so\n switch_model can find their base_url from config\n- Models not in the /v1/models probe but present in the user's saved\n provider config are accepted with a warning instead of rejected\n- CLI /model and TUI /model both pass user_providers/custom_providers\n to switch_model so the config model list is available for validation\n\nCloses #15088\n","name":"fix/model-switch-custom-endpoint","oid":"0d3d2a26316baae7763b29e44dfb33cd914a6fe3"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T10:06:46.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T10:06:46.000000+05:30"},"message":"fix: restore accidentally deleted websocket close code assertion\n","name":"fix/web-env-validation","oid":"9aed1b2fe75d01bf71bd754a863da14fc153b9d7"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T10:05:44.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T10:05:44.000000+05:30"},"message":"chore: add poruru-code to AUTHOR_MAP\n","name":"fix/hindsight-blank-nuke","oid":"2246c81d0d7fc540f6c96b73c3bb4815938d1bae"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T09:56:29.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T09:59:11.000000+05:30"},"message":"fix(setup): skip AUXILIARY_VISION_MODEL write when input is blank\n\nBlank input at the non-OpenAI vision model prompt was unconditionally\nwritten to .env, overwriting any existing custom model.\n","name":"fix/setup-vision-blank","oid":"5eaceb82afe96dac463f9ac6e7deb5862b8b6410"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T00:11:26.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T06:32:39.000000+05:30"},"message":"feat(session): inject Discord IDs block when discord tool is loaded\n\nWhen DISCORD_BOT_TOKEN is set — meaning the discord tool actually\nloads — emit a dedicated IDs block in the session context prompt so\nthe agent can call ``fetch_messages``, ``pin_message``, etc. with\nreal identifiers instead of probing.\n\nCurrently only ``thread_id`` was exposed as a raw ID (via the\n``description`` string). The agent in a Discord thread had to guess\nthat the thread ID doubles as a channel ID for the REST API (it\ndoes), and it had no way to reference the parent channel, the guild,\nor the triggering message at all.\n\nThe block adapts to context:\n\n - Thread: guild / parent channel / thread / message\n - Channel: guild / channel / message\n - (DM has no guild/channel IDs worth listing; only message)\n\nDiscord isn't in _PII_SAFE_PLATFORMS, so IDs ship unredacted.\n","name":"sid/discord-context-injection","oid":"d755601d2679948cb432223ece1775a1f3c556d3"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-24T15:54:03.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T06:32:33.000000+05:30"},"message":"feat(feishu): wire feishu doc/drive tools into hermes-feishu composite\n\nThe feishu_doc and feishu_drive tools were registered in the tool\nregistry but never added to the hermes-feishu composite toolset.\nThe pipeline fix from the prior commit now recovers them automatically\nonce they are in the composite.\n","name":"sid/discord-tool-split","oid":"0d2aa2b6b45f8a6c6c41e36183cb5921f92e7672"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-24T15:55:51.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T06:23:17.000000+05:30"},"message":"fix(tools): normalize numeric entries and clear stale no_mcp in _save_platform_tools\n\nYAML parses bare numeric toolset names (e.g. 12306:) as int, causing\nTypeError in sorted() since the read path normalizes to str but the\nsave path did not.\n\nThe no_mcp sentinel was preserved in existing entries even when the\nuser re-enabled MCP servers, causing MCP to stay silently disabled.\n","name":"sid/fix-tools-config-save","oid":"97a4018dfcdd7daffae944388aed21cf40c84004"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T00:11:26.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-25T05:43:23.000000+05:30"},"message":"feat(session): inject Discord IDs block when discord tool is loaded\n\nWhen DISCORD_BOT_TOKEN is set — meaning the discord tool actually\nloads — emit a dedicated IDs block in the session context prompt so\nthe agent can call ``fetch_messages``, ``pin_message``, etc. with\nreal identifiers instead of probing.\n\nCurrently only ``thread_id`` was exposed as a raw ID (via the\n``description`` string). The agent in a Discord thread had to guess\nthat the thread ID doubles as a channel ID for the REST API (it\ndoes), and it had no way to reference the parent channel, the guild,\nor the triggering message at all.\n\nThe block adapts to context:\n\n - Thread: guild / parent channel / thread / message\n - Channel: guild / channel / message\n - (DM has no guild/channel IDs worth listing; only message)\n\nDiscord isn't in _PII_SAFE_PLATFORMS, so IDs ship unredacted.\n","name":"sid/fix-platform-tool-loading","oid":"7efd91d4b465c26b151b3f5991727ec23e56b875"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-24T16:44:25.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-24T16:44:25.000000-07:00"},"message":"chore(release): add amanning3390 to AUTHOR_MAP\n\nFollow-up for cherry-picked PR #15203 (minimax-oauth provider).\n","name":"hermes/hermes-6c37b3dd","oid":"3812fd81740416b3c7e5696e194e42d8cc2d9dd7"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-24T16:42:43.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-24T16:42:43.000000-07:00"},"message":"rename: shop -> shopify (commerce/shopify)\n\nRenames the skill directory and frontmatter name so the slash command\nbecomes /shopify rather than /shop — clearer attribution and avoids\ncollision with any future generic 'shop' skill.\n","name":"hermes/hermes-9d07c44f","oid":"2d9cc444144ca09009150346de69eb9533f421c5"},{"author":{"email":"ben@nousresearch.com","name":"Ben","when":"2026-04-24T15:29:56.000000+10:00"},"committer":{"email":"ben@nousresearch.com","name":"Ben","when":"2026-04-24T15:29:56.000000+10:00"},"message":"[verified] feat(nous): drive model picker from Portal recommended-models endpoint\n\nReplace the hardcoded _PROVIDER_MODELS[\"nous\"] catalog (~29 entries that\nhad to be updated manually on every Portal model release) with a live\nfetch from /api/nous/recommended-models, keyed off the user's free/paid\ntier. The Portal is now the single source of truth — adding or removing\na Nous model no longer requires a Hermes release.\n\n## What changes\n\nhermes_cli/models.py\n - Remove the hardcoded \"nous\": [...] list from _PROVIDER_MODELS.\n - Add get_nous_recommended_catalog(): reuses the existing 10-minute\n TTL cache in fetch_nous_recommended_models() (no extra HTTP per\n call; shares the cache with the aux/vision model helper). Selects\n freeRecommendedModels vs paidRecommendedModels based on\n check_nous_free_tier(), preserves server-specified ordering\n (the endpoint already orders each array by \"position\"), and does\n case-insensitive dedup.\n - Add _nous_catalog(): exception-safe wrapper returning [] on any\n failure, so callers treat Portal unavailability as \"no catalog\"\n rather than a crash.\n - Rewire provider_model_ids(\"nous\"): Portal recommended-models is\n now primary; the inference /models endpoint stays as a secondary\n live fallback for offline/misconfigured-portal resilience.\n - get_default_model_for_provider(\"nous\") and detect_provider_for_model()\n now route through _nous_catalog() instead of the removed dict key.\n\nhermes_cli/auth.py\n - _login_nous() swapped _PROVIDER_MODELS.get(\"nous\", []) → _nous_catalog().\n\nhermes_cli/main.py\n - /model command nous branch: same swap.\n\n## Design notes\n\n- Free-tier detection happens inside the helper, so single call sites\n don't have to plumb the tier bool around.\n- On tier-detection exception, defaults to paid — matches the existing\n convention (never block paying users).\n- The _AGGREGATORS gate in detect_provider_for_model()'s cross-provider\n match loop already skipped \"nous\" when it was in _PROVIDER_MODELS,\n so removing the key changes nothing in that loop.\n- partition_nous_models_by_tier() is kept in the call sites; it becomes\n mostly a no-op on the server-tier-filtered list but preserves the\n \"upgrade at {portal}\" messaging for free-tier users with no free\n models available.\n\n## Tests\n\ntests/hermes_cli/test_nous_recommended_models.py (new, 22 tests):\n - Server-order preservation\n - Free vs paid routing, auto-detection + exception-defaults-to-paid\n - Empty / missing-field / malformed entries → []\n - Case-insensitive dedup preserving first-seen casing\n - provider_model_ids(\"nous\") rewiring: Portal-first, inference fallback,\n force_refresh propagation, exception-falls-through\n - _PROVIDER_MODELS[\"nous\"] is absent\n - get_default_model_for_provider(\"nous\") Portal-driven,\n non-nous providers unaffected\n - detect_provider_for_model() bare-name + current-provider paths\n - _nous_catalog() swallows exceptions → []\n - curated_models_for_provider(\"nous\") routes through Portal\n\ntests/hermes_cli/test_auth_nous_provider.py:\n - Add get_nous_recommended_catalog stub to _patch_login_internals so\n the login flow has models to present without a live network call.\n\n## Verification\n\nscripts/run_tests.sh tests/hermes_cli/ tests/test_empty_model_fallback.py\n tests/acp/test_server.py tests/test_tui_gateway_server.py\n tests/agent/test_bedrock_integration.py\n → 2752 passed. The 4 remaining failures + 1 collection race are\n pre-existing on main (unrelated — skills filtering, tip length,\n Linux stdlib ssl quirk, xdist race), confirmed via git-stash diff.\n","name":"alice/nous-portal-recommended-models","oid":"71db091868e03fb8d14f66adf37b7e9a06389daf"},{"author":{"email":"arilotter@gmail.com","name":"Ari Lotter","when":"2026-04-11T20:51:34.000000-04:00"},"committer":{"email":"arilotter@gmail.com","name":"Ari Lotter","when":"2026-04-24T00:07:23.000000-04:00"},"message":"feat: add fast-path setup for nous account\n\nadds a nous account specific fast flow & autolaunches into chat if\ngateway isn't set up\n","name":"onboarding","oid":"2f230b5ad988e4bcd83b7fef0c0cd250519d7121"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-24T07:36:03.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-24T07:36:03.000000+05:30"},"message":"Fix ACP module path documentation from acp_adapter to hermes_agent.acp\n","name":"sid/restructure-pr1-acp","oid":"b49bd7b93d1c2b8126dc87f9181487dcb23df7e1"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-19T17:29:51.000000-05:00"},"committer":{"email":"pickett.austin@gmail.com","name":"Austin Pickett","when":"2026-04-23T21:11:04.000000-04:00"},"message":"feat(tui-gateway): WebSocket transport + /chat web UI, wire-compatible with Ink\n\nExtracts the JSON-RPC transport from stdio into an abstraction so the same\ndispatcher drives Ink over stdio AND browser/iOS clients over WebSocket\nwithout duplicating handler logic. Adds a Chat page to the existing web\ndashboard that exercises the full surface — streaming, tool calls, slash\ncommands, model picker, session resume.\n\nBackend\n-------\n* tui_gateway/transport.py — Transport protocol + contextvar binding + the\n module-level StdioTransport. Stream is resolved through a callback so\n tests that monkeypatch `_real_stdout` keep working.\n* tui_gateway/server.py — write_json and dispatch are now transport-aware.\n Backward compatible: no transport bound = legacy stdio path, so entry.py\n (Ink's stdio entrypoint) is unchanged externally.\n* tui_gateway/ws.py — WSTransport + handle_ws coroutine. Safe to call from\n any thread: detects loop-thread deadlock and fire-and-forget schedules\n when needed, blocking run_coroutine_threadsafe + future.result otherwise.\n* hermes_cli/web_server.py — mounts /api/ws on the existing FastAPI app,\n gated by the same ephemeral session token used for REST. Adds\n HERMES_DASHBOARD_DEV_TOKEN env override so Vite HMR dev can share the\n token with the backend.\n\nFrontend\n--------\n* web/src/lib/gatewayClient.ts — browser WebSocket JSON-RPC client that\n mirrors ui-tui/src/gatewayClient.ts.\n* web/src/lib/slashExec.ts — slash command pipeline (slash.exec with\n command.dispatch fallback + exec/plugin/alias/skill/send directive\n handling), mirrors ui-tui/src/app/createSlashHandler.ts.\n* web/src/pages/ChatPage.tsx — transcript + composer driven entirely by\n the WS.\n* web/src/components/SlashPopover.tsx — autocomplete popover above the\n composer, debounced complete.slash.\n* web/src/components/ModelPickerDialog.tsx — two-stage provider/model\n picker; confirms by emitting /model through the slash pipeline.\n* web/src/components/ToolCall.tsx — expandable tool call row (Ink-style\n chevron + context + summary/error/diff).\n* web/src/App.tsx — logo links to /, Chat entry added to nav.\n* web/src/pages/SessionsPage.tsx — every session row gets an Open-in-chat\n button that navigates to /chat?resume= (uses session.resume).\n* web/vite.config.ts — /api proxy configured with ws: true so WebSocket\n upgrades forward in dev mode; injectDevToken plugin reads\n HERMES_DASHBOARD_DEV_TOKEN and injects it into the served index.html so\n Vite HMR can authenticate against FastAPI without a separate flow.\n\nTests\n-----\ntests/hermes_cli/test_web_server.py picks up three new classes:\n\n* TestTuiGatewayWebSocket — handshake, auth rejection, parse errors,\n unknown methods, inline + pool handler round-trips, session event\n routing, disconnect cleanup.\n* TestTuiGatewayTransportParity — byte-identical envelopes for the same\n RPC over stdio vs WS (unknown method, inline handler, error envelope,\n explicit stdio transport).\n* TestTuiGatewayE2EAnyPort — scripted multi-RPC conversation driven\n identically via handle_request and via WebSocket; order + shape must\n match. This is the \"hermes --tui in any port\" check.\n\nExisting tests under tests/tui_gateway/ and tests/test_tui_gateway_server.py\nall still pass unchanged — backward compat preserved.\n\nTry it\n------\n hermes dashboard # builds web, serves on :9119, click Chat\n\nDev with HMR:\n\n export HERMES_DASHBOARD_DEV_TOKEN=\"dev-\\$(openssl rand -hex 16)\"\n hermes dashboard --no-open\n cd web && npm run dev # :5173, /api + /api/ws proxied to :9119\n\nfix(chat): insert tool rows before the streaming assistant message\n\nTranscript used to read \"user → empty assistant bubble → tool → bubble\nfilling in\", which is disorienting: the streaming cursor sits at the top\nwhile the \"work\" rows appear below it chronologically.\n\nNow tool.start inserts the row just before the current streaming\nassistant message, so the order reads \"user → tools → final message\".\nIf no streaming assistant exists yet (rare), tools still append at the\nend; tool.progress / tool.complete match by id regardless of position.\n\nfix(web-chat): font, composer, streaming caret + port GoodVibesHeart\n\n- ChatPage root opts out of App's `font-mondwest uppercase` (dashboard\n chrome style) — adds `font-courier normal-case` so transcript prose is\n readable mono mixed-case instead of pixel-display caps.\n- Composer: textarea + send button wrapped as one bordered unit with\n `focus-within` ring; `font-sans` dropped (it mapped to `Collapse`\n display). Heights stretch together via `items-stretch`; button is a\n flush cap with `border-l` divider.\n- Streaming caret no longer wraps to a new line when the assistant\n renders a block element. Markdown now takes a `streaming` prop and\n injects the caret inside the last block (paragraph, list item, code)\n so it hugs the trailing character. Caret sized in em units.\n- EmptyState gets a blinking caret + shortcut chips.\n- Port ui-tui's GoodVibesHeart easter egg to the web: typing \"thanks\" /\n \"ty\" / \"ily\" / \"good bot\" flashes a Lucide heart next to the\n connection badge (same regex, same 650ms beat, same palette as\n ui-tui/src/app/useMainApp.ts).\n","name":"bb/tui-web-chat","oid":"25ba6783b8eeca3936e10e31c5b5ac9e46a06b09"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-23T17:05:10.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-23T17:05:10.000000-07:00"},"message":"feat(config): make tool output truncation limits configurable\n\nPort from anomalyco/opencode#23770: expose a new `tool_output` config\nsection so users can tune the hardcoded truncation caps that apply to\nterminal output and read_file pagination.\n\nThree knobs under `tool_output`:\n- max_bytes (default 50_000) — terminal stdout/stderr cap\n- max_lines (default 2000) — read_file pagination cap\n- max_line_length (default 2000) — per-line cap in line-numbered view\n\nAll three keep their existing hardcoded values as defaults, so behaviour\nis unchanged when the section is absent. Power users on big-context\nmodels can raise them; small-context local models can lower them.\n\nImplementation:\n- New `tools/tool_output_limits.py` reads the section with defensive\n fallback (missing/invalid values → defaults, never raises).\n- `tools/terminal_tool.py` MAX_OUTPUT_CHARS now comes from\n get_max_bytes().\n- `tools/file_operations.py` normalize_read_pagination() and\n _add_line_numbers() now pull the limits at call time.\n- `hermes_cli/config.py` DEFAULT_CONFIG gains the `tool_output` section\n so `hermes setup` writes defaults into fresh configs.\n- Docs page `user-guide/configuration.md` gains a \"Tool Output\n Truncation Limits\" section with large-context and small-context\n example configs.\n\nTests (18 new in tests/tools/test_tool_output_limits.py):\n- Default resolution with missing / malformed / non-dict config.\n- Full and partial user overrides.\n- Coercion of bad values (None, negative, wrong type, str int).\n- Shortcut accessors delegate correctly.\n- DEFAULT_CONFIG exposes the section with the right defaults.\n- Integration: normalize_read_pagination clamps to the configured\n max_lines.\n","name":"opencode-port/configurable-tool-output-limits","oid":"f2b1b3f1a3ff28ebc39b46cdc02e5febe872281b"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-23T16:44:24.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-23T16:44:24.000000-07:00"},"message":"feat(computer-use): cua-driver backend, universal any-model schema\n\nBackground macOS desktop control via cua-driver MCP — does NOT steal the\nuser's cursor or keyboard focus, works with any tool-capable model.\n\nReplaces the Anthropic-native `computer_20251124` approach from the\nabandoned #4562 with a generic OpenAI function-calling schema plus SOM\n(set-of-mark) captures so Claude, GPT, Gemini, and open models can all\ndrive the desktop via numbered element indices.\n\n## What this adds\n\n- `tools/computer_use/` package — swappable ComputerUseBackend ABC +\n CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).\n- Universal `computer_use` tool with one schema for all providers.\n Actions: capture (som/vision/ax), click, double_click, right_click,\n middle_click, drag, scroll, type, key, wait, list_apps, focus_app.\n- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style\n `content: [text, image_url]` parts) that flows through\n handle_function_call into the tool message. Anthropic adapter converts\n into native `tool_result` image blocks; OpenAI-compatible providers\n get the parts list directly.\n- Image eviction in convert_messages_to_anthropic: only the 3 most\n recent screenshots carry real image data; older ones become text\n placeholders to cap per-turn token cost.\n- Context compressor image pruning: old multimodal tool results have\n their image parts stripped instead of being skipped.\n- Image-aware token estimation: each image counts as a flat 1500 tokens\n instead of its base64 char length (~1MB would have registered as\n ~250K tokens before).\n- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset\n is active.\n- Session DB persistence strips base64 from multimodal tool messages.\n- Trajectory saver normalises multimodal messages to text-only.\n- `hermes tools` post-setup installs cua-driver via the upstream script\n and prints permission-grant instructions.\n- CLI approval callback wired so destructive computer_use actions go\n through the same prompt_toolkit approval dialog as terminal commands.\n- Hard safety guards at the tool level: blocked type patterns\n (curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,\n force delete, lock screen, log out).\n- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)\n workflow guide.\n- Docs: `user-guide/features/computer-use.md` plus reference catalog\n entries.\n\n## Tests\n\n44 new tests in tests/tools/test_computer_use.py covering schema\nshape (universal, not Anthropic-native), dispatch routing, safety\nguards, multimodal envelope, Anthropic adapter conversion, screenshot\neviction, context compressor pruning, image-aware token estimation,\nrun_agent helpers, and universality guarantees.\n\n469/469 pass across tests/tools/test_computer_use.py + the affected\nagent/ test suites.\n\n## Not in this PR\n\n- `model_tools.py` provider-gating: the tool is available to every\n provider. Providers without multi-part tool message support will see\n text-only tool results (graceful degradation via `text_summary`).\n- Anthropic server-side `clear_tool_uses_20250919` — deferred;\n client-side eviction + compressor pruning cover the same cost ceiling\n without a beta header.\n\n## Caveats\n\n- macOS only. cua-driver uses private SkyLight SPIs\n (SLEventPostToPid, SLPSPostEventRecordTo,\n _AXObserverAddNotificationAndCheckRemote) that can break on any macOS\n update. Pin with HERMES_CUA_DRIVER_VERSION.\n- Requires Accessibility + Screen Recording permissions — the post-setup\n prints the Settings path.\n\nSupersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-\nnative schema). Credit @0xbyt4 for the original #3816 groundwork whose\ncontext/eviction/token design is preserved here in generic form.\n","name":"hermes/hermes-34b3f52d","oid":"b07791db0508f92625dfc9e75f20c331cc7bb528"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-23T16:53:25.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-23T17:41:25.000000+05:30"},"message":"refactor: remove config TypedDicts and fix ImportError propagation in clipboard\n\nRemove 44 TypedDict classes from config.py — they were already stale\n(11 missing keys) and load_config() still returns Dict[str, Any], so\nthey provided zero type-checking value. Keep the int() coercions and\nDict[str, Any] annotations which are real fixes.\n\nFix _wayland_save() swallowing ImportError at DEBUG level by adding\nan explicit except ImportError: raise before the broad except Exception.\n","name":"sid/types-and-lints","oid":"415043315f2a7e66f2cc5f3f03f043237019b445"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-23T15:26:41.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-23T15:26:41.000000+05:30"},"message":"chore(restructure): add .git-blame-ignore-revs for restructure commits\n\nTells git blame (and GitHub) to skip the git-mv and import-rewrite\ncommits so blame shows the original author.\n","name":"sid/foundational-restructure","oid":"987962f4531ff879bc4fb16aa0b72537e50aaf78"},{"author":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-22T21:42:14.000000-04:00"},"committer":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-22T21:42:14.000000-04:00"},"message":"Merge remote-tracking branch 'origin/main' into feat/dashboard-chat\n","name":"feat/dashboard-chat","oid":"1cd2b280fdbdb1a1c579795c1a6d2fadf663e410"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-22T17:07:12.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-22T17:07:12.000000-07:00"},"message":"fix(error_classifier): retry mid-stream SSL/TLS alert errors as transport\n\nMid-stream SSL alerts (bad_record_mac, tls_alert_internal_error, handshake\nfailures) previously fell through the classifier pipeline to the 'unknown'\nbucket because:\n\n - ssl.SSLError type names weren't in _TRANSPORT_ERROR_TYPES (the\n isinstance(OSError) catch picks up some but not all SDK-wrapped forms)\n - the message-pattern list had no SSL alert substrings\n\nThe 'unknown' bucket is still retryable, but: (a) logs tell the user\n'unknown' instead of identifying the cause, (b) it bypasses the\ntransport-specific backoff/fallback logic, and (c) if the SSL error\nhappens on a large session with a generic 'connection closed' wrapper,\nthe existing disconnect-on-large-session heuristic would incorrectly\ntrigger context compression — expensive, and never fixes a transport\nhiccup.\n\nChanges:\n - Add ssl.SSLError and its subclass type names to _TRANSPORT_ERROR_TYPES\n - New _SSL_TRANSIENT_PATTERNS list (separate from _SERVER_DISCONNECT_PATTERNS\n so SSL alerts route to timeout, not context_overflow+compress)\n - New step 5 in the classifier pipeline: SSL pattern check runs BEFORE\n the disconnect check to pre-empt the large-session-compress path\n\nPatterns cover both space-separated ('ssl alert', 'bad record mac')\nand underscore-separated ('ERR_SSL_SSL/TLS_ALERT_BAD_RECORD_MAC')\nforms. This is load-bearing because OpenSSL 3.x changed the error-code\nseparator from underscore to slash (e.g. SSLV3_ALERT_BAD_RECORD_MAC →\nSSL/TLS_ALERT_BAD_RECORD_MAC) and will likely churn again — matching on\nstable alert reason substrings survives future format changes.\n\nTests (8 new):\n - BAD_RECORD_MAC in Python ssl.c format\n - OpenSSL 3.x underscore format\n - TLSV1_ALERT_INTERNAL_ERROR\n - ssl handshake failure\n - [SSL: ...] prefix fallback\n - Real ssl.SSLError instance\n - REGRESSION GUARD: SSL on large session does NOT compress\n - REGRESSION GUARD: plain disconnect on large session STILL compresses\n","name":"gemini-cli-port/ssl-retry-patterns","oid":"b40b6ec720df05786b865579d3a094efe9e481a2"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-22T17:03:35.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-22T17:03:35.000000-07:00"},"message":"fix(usage): read top-level Anthropic cache fields from OAI-compatible proxies\n\nPort from cline/cline#10266.\n\nWhen OpenAI-compatible proxies (OpenRouter, Vercel AI Gateway, Cline)\nroute Claude models, they sometimes surface the Anthropic-native cache\ncounters (`cache_read_input_tokens`, `cache_creation_input_tokens`) at\nthe top level of the `usage` object instead of nesting them inside\n`prompt_tokens_details`. Our chat-completions branch of\n`normalize_usage()` only read the nested `prompt_tokens_details` fields,\nso those responses:\n\n- reported `cache_write_tokens = 0` even when the model actually did a\n prompt-cache write,\n- reported only some of the cache-read tokens when the proxy exposed them\n top-level only,\n- overstated `input_tokens` by the missed cache-write amount, which in\n turn made cost estimation and the status-bar cache-hit percentage wrong\n for Claude traffic going through these gateways.\n\nNow the chat-completions branch tries the OpenAI-standard\n`prompt_tokens_details` first and falls back to the top-level\nAnthropic-shape fields only if the nested values are absent/zero. The\nAnthropic and Codex Responses branches are unchanged.\n\nRegression guards added for three shapes: top-level write + nested read,\ntop-level-only, and both-present (nested wins).\n","name":"cline-port/anthropic-cache-tokens-top-level","oid":"a369987443013afb29581ed272900a1f01451345"},{"author":{"email":"dmayhem93@gmail.com","name":"dmahan93","when":"2026-04-22T14:57:57.000000-05:00"},"committer":{"email":"dmayhem93@gmail.com","name":"dmahan93","when":"2026-04-22T14:57:57.000000-05:00"},"message":"final changes from successful run\n","name":"nemo-gym-changes","oid":"be43bee11af0760f395b6139c9031d8656749a67"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T23:39:02.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T23:49:50.000000+05:30"},"message":"fix(state): replace version-gated migrations with declarative column reconciliation\n\nThe linear migration chain (if current_version < N: ALTER TABLE ADD COLUMN)\nbroke when commit a7d78d3b inserted a new v7 migration (reasoning_content)\nand renumbered the old v7 (api_call_count) to v8. Users who updated between\nApr 15–22 already had schema_version=7 from the old numbering, so the new\nv7 block was skipped entirely — reasoning_content was never created, causing\n'sqlite3.OperationalError: no such column: reasoning_content' on /continue.\n\nRoot cause: two independent sources of truth (SCHEMA_SQL for new DBs vs a\nmanually-maintained version chain for existing DBs) that could desync when\nmigrations were reordered.\n\nFix: replace the 8-block version chain with _reconcile_columns(), which on\nevery startup diffs live table columns (PRAGMA table_info) against the\ncolumns declared in SCHEMA_SQL and ADDs any that are missing. This follows\nthe Beets/sqlite-utils pattern used across mature SQLite projects — the\nCREATE TABLE definition becomes the single source of truth for schema.\n\nAdding a new column is now a single change: add it to SCHEMA_SQL. No\nmigration block to write, no version number to bump, no possibility of\nskipped columns from reordering.\n\nThe schema_version table is retained for future data migrations that\ncannot be handled declaratively (row transforms), but zero version-gated\nADD COLUMN blocks remain.\n","name":"fix/schema-reconciliation","oid":"ccd4116635fd5031fbb9ce53ad0db904e1bf0fef"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T23:05:42.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T23:05:42.000000+05:30"},"message":"perf(browser): upgrade agent-browser 0.13 -> 0.26, wire daemon idle timeout\n\nUpgrades agent-browser from 0.13.0 to 0.26.0, picking up 13 releases of\ndaemon reliability fixes:\n\n- Daemon hang on Linux from waitpid(-1) race in SIGCHLD handler (#1098)\n- Chrome killed after ~10s idle due to PR_SET_PDEATHSIG thread tracking (#1157)\n- Orphaned Chrome processes via process-group kill on shutdown (#1137)\n- Stale daemon after upgrade via .version sidecar and auto-restart (#1134)\n- Idle timeout not firing (sleep future recreated each loop) (#1110)\n- Navigation hanging on lifecycle events that never fire (#1059, #1092)\n- CDP attach hang on Chrome 144+ (#1133)\n- Windows daemon TCP bind with Hyper-V port conflicts (#1041)\n- Shadow DOM traversal in accessibility tree snapshots\n- doctor command for user self-diagnosis\n\nAlso wires AGENT_BROWSER_IDLE_TIMEOUT_MS into the browser subprocess\nenvironment so the daemon self-terminates after our configured inactivity\ntimeout (default 300s). This is the daemon-side counterpart to the\nPython-side inactivity reaper — the daemon kills itself and its Chrome\nchildren when no commands arrive, preventing orphan accumulation even\nwhen the Python process dies without running atexit handlers.\n\nAddresses #7343 (daemon socket hangs, shadow DOM) and #13793 (orphan\naccumulation from force-killed sessions).\n","name":"fix/upgrade-agent-browser-0.26","oid":"81fe89aca2cbde70bf6a80070c3cbca8a4f2df35"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T21:32:51.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T22:50:54.000000+05:30"},"message":"feat(security): add global toggle to allow private/internal URL resolution\n\nOn networks using OpenWrt routers, corporate proxies, or VPNs that\nresolve external domains to private IP ranges (198.18.0.0/15, 100.64.x,\netc.), Hermes blocks ALL outbound requests because is_safe_url() treats\nany private IP as an SSRF attack vector.\n\nThis adds a global toggle that disables private-IP blocking across all\n23 call sites (web_tools, vision_tools, browser_tool, and 13 gateway\nplatform adapters) from a single config key.\n\nSecurity guarantee: cloud metadata endpoints are ALWAYS blocked\nregardless of the toggle:\n- Hostnames: metadata.google.internal, metadata.goog\n- IPs: 169.254.169.254, fd00:ec2::254\n\nThree ways to enable (priority order):\n1. HERMES_ALLOW_PRIVATE_URLS=true env var\n2. security.allow_private_urls: true in config.yaml\n3. browser.allow_private_urls: true (legacy, now promotes globally)\n\nFiles changed:\n- tools/url_safety.py: _global_allow_private_urls() with cached config\n read, _ALWAYS_BLOCKED_IPS for metadata endpoints, is_safe_url() checks\n the toggle after blocking metadata hostnames\n- hermes_cli/config.py: security.allow_private_urls in DEFAULT_CONFIG\n- tests/tools/test_url_safety.py: 32 new tests covering toggle, config\n fallback, caching, and security invariants (metadata always blocked)\n","name":"feat/ssrf-allow-private-urls","oid":"762e1073a8c27ce48bb7afe3a78b748b4621b0e6"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T22:41:38.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T22:42:39.000000+05:30"},"message":"fix: follow-up for salvaged PR #8952\n\n- Rename provider_contracts.py -> volcengine_byteplus.py for explicitness\n- Consolidate duplicate host-to-provider mappings: provider_for_base_url()\n now uses the canonical _URL_TO_PROVIDER from model_metadata.py instead of\n maintaining a separate 20-entry dict\n- Add volcengine/byteplus to runtime_provider.py model-dependent base URL\n resolution (kimi-style special case) so manually-edited configs resolve\n the coding-plan base URL correctly\n- Remove volcengine/byteplus from _API_KEY_PROVIDER_AUX_MODELS — the\n main-model-first design in _resolve_auto() handles these providers\n already; entries were dead code in the normal flow\n- Add VOLCENGINE_API_KEY and BYTEPLUS_API_KEY to OPTIONAL_ENV_VARS in\n config.py so they appear in hermes setup\n- Update docs: environment-variables.md, fallback-providers.md,\n configuration.md\n","name":"feat/volcengine-byteplus","oid":"8aaefec231576b1fde5f4a202e194f09bafd261b"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T11:01:28.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T14:30:09.000000+05:30"},"message":"refactor: unify transport dispatch + collapse normalize shims\n\nConsolidate 4 per-transport lazy singleton helpers (_get_anthropic_transport,\n_get_codex_transport, _get_chat_completions_transport, _get_bedrock_transport)\ninto one generic _get_transport(api_mode) with a shared dict cache.\n\nCollapse the 65-line main normalize block (3 api_mode branches, each with\nits own SimpleNamespace shim) into 7 lines: one _get_transport() call +\none _nr_to_assistant_message() shared shim. The shim extracts provider_data\nfields (codex_reasoning_items, reasoning_details, call_id, response_item_id)\ninto the SimpleNamespace shape downstream code expects.\n\nWire chat_completions and bedrock_converse normalize through their transports\nfor the first time — these were previously falling into the raw\nresponse.choices[0].message else branch.\n\nRemove 8 dead codex adapter imports that have zero callers after PRs 1-6.\n\nTransport lifecycle improvements:\n- Eagerly warm transport cache at __init__ (surfaces import errors early)\n- Invalidate transport cache on api_mode change (switch_model, fallback\n activation, fallback restore, transport recovery) — prevents stale\n transport after mid-session provider switch\n\nrun_agent.py: -32 net lines (11,988 -> 11,956).\n\nPR 7 of the provider transport refactor.\n","name":"refactor/unify-transport-dispatch","oid":"44596731c82c090fc6eda7244b697d58f69c65b6"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T10:27:46.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T10:39:08.000000+05:30"},"message":"feat(gateway): add require_mention_channels for per-channel mention overrides\n\nAdds a new `require_mention_channels` config key (and corresponding env\nvars) across all 7 gateway platforms that have mention-gating. This is\nthe inverse of `free_response_channels` — channels listed here always\nrequire @mention even when the global `require_mention` setting is false.\n\nUse case: a user runs multiple channels, most with a single agent where\nno mention is needed, but a few 'agent group' channels where mentions\nmake sense to avoid noise. Previously this required require_mention=true\nglobally and listing every non-group channel in free_response_channels.\n\nPriority logic (highest wins):\n1. DMs → always respond\n2. Channel in free_response_channels → never require mention\n3. Channel in require_mention_channels → always require mention\n4. Global require_mention setting → fallback\n\nAlso refactors:\n- Mattermost: inline os.getenv() → proper helper methods matching\n Discord/Slack pattern, adds config.yaml support via config.extra\n- Matrix: cached __init__ vars → helper methods with config.extra\n support\n- Mattermost config bridging: adds missing YAML→env bridging in\n gateway/config.py (was completely absent)\n\nNew env vars:\n- DISCORD_REQUIRE_MENTION_CHANNELS\n- SLACK_REQUIRE_MENTION_CHANNELS\n- TELEGRAM_REQUIRE_MENTION_CHATS\n- WHATSAPP_REQUIRE_MENTION_CHATS\n- DINGTALK_REQUIRE_MENTION_CHATS\n- MATTERMOST_REQUIRE_MENTION_CHANNELS\n- MATRIX_REQUIRE_MENTION_ROOMS\n\nNaming follows each platform's convention (channels/chats/rooms).\nFully backward compatible — empty by default, no behavior change.\n\nRequested by community member neeldhara on PR #3664.\n","name":"feat/require-mention-channels","oid":"d4178e09776c4ff5711e8afc88d3089b1665c1d1"},{"author":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-21T22:53:45.000000-04:00"},"committer":{"email":"emozilla@nousresearch.com","name":"emozilla","when":"2026-04-21T22:53:45.000000-04:00"},"message":"feat(aux): use Portal /api/nous/recommended-models for auxiliary models\n\nWire the auxiliary client (compaction, vision, session search, web extract)\nto the Nous Portal's curated recommended-models endpoint when running on\nNous Portal, with a TTL-cached fetch that mirrors how we pull /models for\npricing.\n\nhermes_cli/models.py\n - fetch_nous_recommended_models(portal_base_url, force_refresh=False)\n 10-minute TTL cache, keyed per portal URL (staging vs prod don't\n collide). Public endpoint, no auth required. Returns {} on any\n failure so callers always get a dict.\n - get_nous_recommended_aux_model(vision, free_tier=None, ...)\n Tier-aware pick from the payload:\n - Paid tier → paidRecommended{Vision,Compaction}Model, falling back\n to freeRecommended* when the paid field is null (common during\n staged rollouts of new paid models).\n - Free tier → freeRecommended* only, never leaks paid models.\n When free_tier is None, auto-detects via the existing\n check_nous_free_tier() helper (already cached 3 min against\n /api/oauth/account). Detection errors default to paid so we never\n silently downgrade a paying user.\n\nagent/auxiliary_client.py — _try_nous()\n - Replaces the hardcoded xiaomi/mimo free-tier branch with a single call\n to get_nous_recommended_aux_model(vision=vision).\n - Falls back to _NOUS_MODEL (google/gemini-3-flash-preview) when the\n Portal is unreachable or returns a null recommendation.\n - The Portal is now the source of truth for aux model selection; the\n xiaomi allowlist we used to carry is effectively dead.\n\nTests (15 new)\n - tests/hermes_cli/test_models.py::TestNousRecommendedModels\n Fetch caching, per-portal keying, network failure, force_refresh;\n paid-prefers-paid, paid-falls-to-free, free-never-leaks-paid,\n auto-detect, detection-error → paid default, null/blank modelName\n handling.\n - tests/agent/test_auxiliary_client.py::TestNousAuxiliaryRefresh\n _try_nous honors Portal recommendation for text + vision, falls\n back to google/gemini-3-flash-preview on None or exception.\n\nBehavior won't visibly change today — both tier recommendations currently\npoint at google/gemini-3-flash-preview — but the moment the Portal ships\na better paid recommendation, subscribers pick it up within 10 minutes\nwithout a Hermes release.\n","name":"remove-nous-free-allowlist","oid":"338b98161ad08fde4423ace88e542625fc834633"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-21T17:09:49.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-21T17:09:49.000000-07:00"},"message":"feat(cli): add --ignore-user-config and --ignore-rules flags\n\nPort from openai/codex#18646.\n\nAdds two flags to 'hermes chat' that fully isolate a run from user-level\nconfiguration and rules:\n\n* --ignore-user-config: skip ~/.hermes/config.yaml and fall back to\n built-in defaults. Credentials in .env are still loaded so the agent\n can actually call a provider.\n* --ignore-rules: skip auto-injection of AGENTS.md, SOUL.md,\n .cursorrules, and persistent memory (maps to AIAgent(skip_context_files=True,\n skip_memory=True)).\n\nPrimary use cases:\n- Reproducible CI runs that should not pick up developer-local config\n- Third-party integrations (e.g. Chronicle in Codex) that bring their\n own config and don't want user preferences leaking in\n- Bug-report reproduction without the reporter's personal overrides\n- Debugging: bisect 'was it my config?' vs 'real bug' in one command\n\nBoth flags are registered on the parent parser AND the 'chat' subparser\n(with argparse.SUPPRESS on the subparser to avoid overwriting the parent\nvalue when the flag is placed before the subcommand, matching the\nexisting --yolo/--worktree/--pass-session-id pattern).\n\nEnv vars HERMES_IGNORE_USER_CONFIG=1 and HERMES_IGNORE_RULES=1 are set\nby cmd_chat BEFORE 'from cli import main' runs, which is critical\nbecause cli.py evaluates CLI_CONFIG = load_cli_config() at module import\ntime. The cli.py / hermes_cli.config.load_cli_config() function checks\nthe env var and skips ~/.hermes/config.yaml when set.\n\nTests: 11 new tests in tests/hermes_cli/test_ignore_user_config_flags.py\ncovering the env gate, constructor wiring, cmd_chat simulation, and\nargparse flag registration. All pass; existing hermes_cli + cli suites\nunaffected (3005 pass, 2 pre-existing unrelated failures).\n","name":"codex-port/ignore-user-config-flags","oid":"569faf54dec950157c66facb183c7fee8e84ddf6"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-21T17:31:01.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-21T17:31:01.000000-05:00"},"message":"feat(tui): add /mouse [on|off|toggle] runtime slash command\n\nToggle SGR mouse tracking (DEC 1000/1002/1003/1006) at runtime without\nrestart or env-var spelunking. Fix path when a terminal doesn't honor raw\nmode / no-echo and echoes mouse events as visible escape sequences (e.g.\n`<35;111;133M` scrolling up the transcript on every mouse move).\n\n- New `/mouse [on|off|toggle]` slash command (persists via config.set key=mouse\n → display.tui_mouse in ~/.hermes/config.yaml).\n- New hermes-ink export `setAltScreenMouseTracking(enabled)` that writes\n ENABLE/DISABLE bytes and updates the instance flag without re-entering\n the alt-screen — so live toggles are flicker-free.\n- `` mouseTracking prop is frozen at initial value (from\n `HERMES_TUI_DISABLE_MOUSE` env); runtime state lives in `$uiState` and is\n applied via useEffect. Env-var opt-out wins over config so explicit\n HERMES_TUI_DISABLE_MOUSE=1 stays off regardless of persisted state.\n- Server: folds `mouse` into the existing compact/statusbar branch in\n config.set/get, defaulting to on.\n","name":"bb/tui-mouse-toggle","oid":"c275423d0d82e23382156e7e2df403b2826bec67"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T01:33:14.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-22T01:49:49.000000+05:30"},"message":"feat(skills): expand touchdesigner-mcp with GLSL, post-FX, audio, geometry references\n\nAdd 6 new reference files with generic reusable patterns:\n- glsl.md: uniforms, built-in functions, shader templates, Bayer dither\n- postfx.md: bloom, CRT scanlines, chromatic aberration, feedback glow\n- layout-compositor.md: layoutTOP, overTOP grids, panel dividers\n- operator-tips.md: wireframe rendering, feedback TOP setup\n- geometry-comp.md: instancing, POP vs SOP rendering, shape morphing\n- audio-reactive.md: band extraction (audiofilterCHOP), beat detection, MIDI\n\nExpand SKILL.md with:\n- TD 2025 API quirks (connection syntax, GLSL TOP rules, expression gotchas)\n- Trimmed param name table (8 known LLM traps, defers to td_get_par_info)\n- Slider-to-shader wiring (td_execute_python + ParMode.EXPRESSION)\n- Frame capture with run()/delayFrames (TOP.save() timing fix)\n- TD 099 POP vs SOP rendering rules\n- Incremental build strategy for large scripts\n- Remote TD setup (PC over Ethernet)\n- Audio synthesis via CHOPs (LFO-driven envelope pattern)\n\nExpand pitfalls.md (#46-63):\n- Connection syntax, moviefileoutTOP bug, batch frame capture\n- TOP.save() time advancement, feedback masking, incremental builds\n- MCP reconnection after project.load(), TOX reverse-engineering\n- sliderCOMP naming, create() suffix requirement\n- COMP reparenting (copyOPs), expressionCHOP crash\n\nAll content is generic — no session-specific paths, hardware, aesthetics,\nor param-name-only entries (those belong in td_get_par_info).\nBumps version 1.0.0 → 2.0.0.\n","name":"feat/td-skill-update","oid":"be5a2ee5d3492dd0b96c2009f07f8eb6658b932d"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-21T15:53:46.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-21T16:15:52.000000+05:30"},"message":"feat: add BedrockTransport + wire all Bedrock transport paths\n\nAdd BedrockTransport wrapping agent/bedrock_adapter.py behind the\nProviderTransport ABC. Fourth and final transport.\n\nWire ALL transport methods to production paths in run_agent.py:\n- build_kwargs: _build_api_kwargs bedrock branch (L6713)\n- normalize_response: main normalize loop, new bedrock_converse branch\n (handles both raw boto3 dicts and already-normalized SimpleNamespace)\n- validate_response: response validation, new bedrock_converse branch\n- finish_reason: new bedrock_converse branch in finish_reason extraction\n\nThe truncation path (L9588) intentionally groups bedrock with\nchat_completions — both have the same response.choices shape because\nnormalize_converse_response runs at the dispatch site.\n\n17 new tests. 231 bedrock/converse/transport tests pass (0 failures).\n\nPR 6 of the provider transport refactor.\n","name":"feat/bedrock-transport","oid":"f56c373a279e4a61c761d7a2f17e6e6547594e32"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-21T15:15:19.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-21T16:00:41.000000+05:30"},"message":"feat: add ChatCompletionsTransport + wire all default paths\n\nAdd ChatCompletionsTransport for the default api_mode used by ~16\nOpenAI-compatible providers (OpenRouter, Nous, NVIDIA, Qwen, Ollama,\nDeepSeek, xAI, custom, etc.).\n\nWire ALL transport methods to production paths in run_agent.py:\n- build_kwargs: extract 210-line else branch with 13 provider-specific\n conditionals (Qwen portal, NVIDIA NIM, Ollama, reasoning, developer\n role swap, provider preferences, max_tokens defaults)\n- validate_response: response.choices validation gate\n- extract_cache_stats: OpenRouter prompt_tokens_details extraction\n- convert_messages: codex field sanitization (identity otherwise)\n- convert_tools: identity (already in OpenAI format)\n- normalize_response: near-identity wrapper returning NormalizedResponse\n\nAgent gathers state (provider detection flags, temperature, preferences)\nand passes as explicit params to transport.build_kwargs().\n\nrun_agent.py: -197 lines in _build_api_kwargs (12,054 -> 11,948).\n\n26 new tests (build_kwargs: 12, validate: 4, normalize: 2, cache: 3,\nconvert: 3, registration: 2). All transport tests pass.\n\nPR 5 of the provider transport refactor.\n","name":"feat/chat-completions-transport","oid":"7570414e59267a273345d5c6dfaea76d1fd0cad5"},{"author":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-21T03:30:19.000000-07:00"},"committer":{"email":"127238744+teknium1@users.noreply.github.com","name":"teknium1","when":"2026-04-21T03:30:19.000000-07:00"},"message":"fix(acp): isolate per-session approval callback via ContextVar\n\nConcurrent ACP sessions in one Hermes process previously shared\ntools.terminal_tool._approval_callback as a module-global, so session B\noverwriting the slot could route session A's dangerous-command prompt\nthrough B's callback (and vice versa). Within a single OS user this was\nUX confusion rather than a cross-principal boundary break, but the\nshared state is genuine concurrency sloppiness worth fixing.\n\nStore the callback (and the sibling sudo password callback) in\nContextVars. Each asyncio task gets its own copy, so per-session\nset_approval_callback calls no longer stomp on each other. ACP's prompt\nhandler now wraps loop.run_in_executor in contextvars.copy_context().run\nso the per-session callback survives the hop into the worker thread —\nasyncio does not propagate contextvars across the executor boundary on\nits own, and this was verified empirically.\n\nRegression tests reproduce the original primitive (two overlapping\nsessions, each asserts it observes its own callback) and document the\nrun_in_executor contextvar contract the ACP fix relies on.\n\nReported by @xeloxa in GHSA-qg5c-hvr5-hjgr.\n","name":"hermes/hermes-bb7c1b2e","oid":"603a68b91846a9413c833496d89805d8e580e1de"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-20T17:04:30.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-20T17:04:30.000000-07:00"},"message":"feat(models): hide OpenRouter models that don't advertise tool support\n\nPort from Kilo-Org/kilocode#9068.\n\nhermes-agent is tool-calling-first — every provider path assumes the\nmodel can invoke tools. Models whose OpenRouter supported_parameters\ndoesn't include 'tools' (e.g. image-only or completion-only models)\ncannot be driven by the agent loop and fail at the first tool call.\n\nFilter them out of fetch_openrouter_models() so they never appear in\nthe model picker (`hermes model`, setup wizard, /model slash command).\n\nPermissive when the field is missing — OpenRouter-compatible gateways\n(Nous Portal, private mirrors, older snapshots) don't always populate\nsupported_parameters. Treat missing as 'unknown → allow' rather than\nsilently emptying the picker on those gateways. Only hide models\nwhose supported_parameters is an explicit list that omits tools.\n\nTests cover: tools present → kept, tools absent → dropped, field\nmissing → kept, malformed non-list → kept, non-dict item → kept,\nempty list → dropped.\n","name":"kilocode-port/filter-non-tool-openrouter-models","oid":"e92d9eb5efa33c9deded5e8a706a1973423b5266"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T23:53:24.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-21T00:00:45.000000+05:30"},"message":"fix(kimi): drop client-side temperature overrides for Kimi/Moonshot models\n\nThe Kimi gateway selects the correct temperature server-side based on the\nactive mode (thinking on → 1.0, thinking off → 0.6). Client-side clamping\nis no longer needed and would conflict if the gateway changes its defaults.\n\nRemoved:\n- _FIXED_TEMPERATURE_MODELS, _KIMI_INSTANT_MODELS, _KIMI_THINKING_MODELS,\n _KIMI_PUBLIC_API_OVERRIDES maps from auxiliary_client.py\n- All Kimi-specific branches in _fixed_temperature_for_model() — the\n function now always returns None (kept for future non-Kimi contracts)\n\nCallers already guard with 'if fixed_temperature is not None:' so the\nchange is transparent — temperature is simply omitted from API calls,\nletting the Kimi gateway use its own defaults.\n\nUpdated tests across 5 files to verify temperature is NOT forced.\n","name":"fix/kimi-drop-temperature","oid":"ed201cce9cd98fe83433e38ea93fbf13bb069862"},{"author":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-20T13:02:32.000000-05:00"},"committer":{"email":"brooklyn.bb.nicholson@gmail.com","name":"Brooklyn Nicholson","when":"2026-04-20T13:02:32.000000-05:00"},"message":"chore: uptick\n","name":"bb/tui-elapsed-lastmsg-8541","oid":"f7587dd445deda6ee51fee8d5c210a1fcd33d672"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T20:13:33.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T20:13:33.000000+05:30"},"message":"feat: add transport types + migrate Anthropic normalize path\n\nAdd agent/transports/types.py with three shared dataclasses:\n- NormalizedResponse: content, tool_calls, finish_reason, reasoning, usage, provider_data\n- ToolCall: id, name, arguments, provider_data (per-tool-call protocol metadata)\n- Usage: prompt_tokens, completion_tokens, total_tokens, cached_tokens\n\nAdd normalize_anthropic_response_v2() to anthropic_adapter.py — wraps the\nexisting v1 function and maps its output to NormalizedResponse. One call site\nin run_agent.py (the main normalize branch) uses v2 with a back-compat shim\nto SimpleNamespace for downstream code.\n\nNo ABC, no registry, no streaming, no client lifecycle. Those land in PR 3\nwith the first concrete transport (AnthropicTransport).\n\n46 new tests:\n- test_types.py: dataclass construction, build_tool_call, map_finish_reason\n- test_anthropic_normalize_v2.py: v1-vs-v2 regression tests (text, tools,\n thinking, mixed, stop reasons, mcp prefix stripping, edge cases)\n\nPart of the provider transport refactor (PR 2 of 9).\n","name":"feat/transport-types","oid":"da032998b4141fd99c1c8fc3d9be5b36a7d77e77"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T16:19:44.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T16:32:43.000000+05:30"},"message":"refactor: extract codex_responses logic into dedicated adapter\n\nMove 10 Responses API format-conversion and normalization functions from\nrun_agent.py into agent/codex_responses_adapter.py. All functions are now\nstateless module-level functions with zero self references.\n\nThe AIAgent methods remain as thin one-line wrappers that delegate to the\nadapter, so all callers (tests, gateway, CLI) are unchanged.\n\nFunctions extracted:\n- _deterministic_call_id: deterministic tool call ID generation\n- _split_responses_tool_id: composite ID splitting\n- _derive_responses_function_call_id: call_ to fc_ prefix conversion\n- _responses_tools: chat completions tool schema → Responses format\n- _chat_messages_to_responses_input: message format conversion\n- _preflight_codex_input_items: input item normalization\n- _preflight_codex_api_kwargs: API kwargs validation/cleaning\n- _extract_responses_message_text: text extraction from response items\n- _extract_responses_reasoning_text: reasoning extraction\n- _normalize_codex_response: full response normalization\n\nThis brings codex_responses in line with anthropic_adapter.py and\nbedrock_adapter.py which already have their own adapter files.\n\nrun_agent.py: 12410 → 11845 lines (-565 net)\n","name":"refactor/extract-codex-adapter","oid":"b277962dcc21968a8967769a899f45495742bfb7"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:35:01.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:35:01.000000-07:00"},"message":"feat(security): URL query param + userinfo + form body redaction\n\nPort from nearai/ironclaw#2529.\n\nHermes already has broad value-shape coverage in agent/redact.py\n(30+ vendor prefixes, JWTs, DB connstrs, etc.) but missed three\nkey-name-based patterns that catch opaque tokens without recognizable\nprefixes:\n\n1. URL query params - OAuth callback codes (?code=...),\n access_token, refresh_token, signature, etc. These are opaque and\n won't match any prefix regex. Now redacted by parameter NAME.\n\n2. URL userinfo (https://user:pass@host) - for non-DB schemes. DB\n schemes were already handled by _DB_CONNSTR_RE.\n\n3. Form-urlencoded body (k=v pairs joined by ampersands) -\n conservative, only triggers on clean pure-form inputs with no\n other text.\n\nSensitive key allowlist matches ironclaw's (exact case-insensitive,\nNOT substring - so token_count and session_id pass through).\n\nTests: +20 new test cases across 3 test classes. All 75 redact tests\npass; gateway/test_pii_redaction and tools/test_browser_secret_exfil\nalso green.\n\nKnown pre-existing limitation: _ENV_ASSIGN_RE greedy match swallows\nwhole all-caps ENV-style names + trailing text when followed by\nanother assignment. Left untouched here (out of scope); URL query\nredaction handles the lowercase case.\n","name":"ironclaw-port/url-and-form-redaction","oid":"b8d00c6f944115f530a896df3f3ff0f952d9f40c"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:22:26.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:22:43.000000-07:00"},"message":"fix(display): strip standalone tool-call XML tags from visible text\n\nPort from openclaw/openclaw#67318. Some open models (notably Gemma\nvariants served via OpenRouter) emit tool calls as XML blocks inside\nassistant content instead of via the structured tool_calls field:\n\n /tmp/x\n {\"name\":\"x\"}\n [{...}]\n\nLeft unstripped, this raw XML leaked to gateway users (Discord, Telegram,\nMatrix, Feishu, Signal, WhatsApp, etc.) and the CLI, since hermes-agent's\nexisting reasoning-tag stripper handled only //\nvariants.\n\nExtend _strip_think_blocks (run_agent.py) and _strip_reasoning_tags\n(cli.py) to cover:\n * , , \n * , \n * ... (Gemma-style)\n\nThe variant is boundary-gated (only strips when the tag sits\nat start-of-line or after sentence punctuation AND carries a name=\"...\"\nattribute) so prose mentions like 'Use declarations in JS'\nare preserved. Dangling with no close is\nintentionally left visible — matches OpenClaw's asymmetry so a truncated\nstreaming tail still reaches the user.\n\nTests: 9 new cases in TestStripThinkBlocks (run_agent) + 9 in new file\ntests/run_agent/test_strip_reasoning_tags_cli.py. Covers Qwen-style\n, Gemma-style , multi-line payloads,\nprose preservation, stray close tags, dangling open tags, and mixed\nreasoning+tool_call content.\n\nNote: this port covers the post-streaming final-text path, which is what\ngateway adapters and CLI display consume. Extending the per-delta stream\nfilter in gateway/stream_consumer.py to hide these tags live as they\nstream is a separate follow-up; for now users may see raw XML briefly\nduring a stream before the final cleaned text replaces it.\n\nRefs: openclaw/openclaw#67318\n","name":"openclaw-port/display-tool-call-tags-strip","oid":"ef1a7b690dc71444cb7f7896c369bfd8637b008a"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:18:23.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:18:23.000000-07:00"},"message":"fix(anthropic): guard max_tokens against non-positive values\n\nPort from openclaw/openclaw#66664. The build_anthropic_kwargs call site\nused 'max_tokens or _get_anthropic_max_output(model)', which correctly\nfalls back when max_tokens is 0 or None (falsy) but lets negative ints\n(-1, -500), fractional floats (0.5, 8192.7), NaN, and infinity leak\nthrough to the Anthropic API. Anthropic rejects these with HTTP 400\n('max_tokens: must be greater than or equal to 1'), turning a local\nconfig error into a surprise mid-conversation failure.\n\nAdd two resolver helpers matching OpenClaw's:\n _resolve_positive_anthropic_max_tokens — returns int(value) only if\n value is a finite positive number; excludes bools, strings, NaN,\n infinity, sub-one positives (floor to 0).\n _resolve_anthropic_messages_max_tokens — prefers a positive requested\n value, else falls back to the model's output ceiling; raises\n ValueError only if no positive budget can be resolved.\n\nThe context-window clamp at the call site (max_tokens > context_length)\nis preserved unchanged — it handles oversized values; the new resolver\nhandles non-positive values. These concerns are now cleanly separated.\n\nTests: 17 new cases covering positive/zero/negative ints, fractional\nfloats (both >1 and <1), NaN, infinity, booleans, strings, None, and\nintegration via build_anthropic_kwargs.\n\nRefs: openclaw/openclaw#66664\n","name":"openclaw-port/anthropic-max-tokens-guard","oid":"b0e398ef7ca9181adaea2d47bf6e1449f713d02f"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:13:42.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T17:13:42.000000-07:00"},"message":"fix(compaction): redact credential-like values from summary pipeline\n\nPort from openclaw/openclaw#67801. The context compressor's summarizer\nprompt instructs the model to preserve specific values (file paths,\ncommands, error messages, etc.) so it can produce concrete handoffs.\nThat instruction also caused API keys, bearer tokens, and env-var\nassignments surfaced through tool output (terminal, read_file, curl -v)\nto be copied verbatim into the persistent summary and re-injected on\nevery subsequent compaction.\n\nApply agent.redact.redact_sensitive_text at three points:\n - serializer output (primary defense)\n - previous-summary re-injection on iterative compaction\n - LLM-returned summary before storage in _previous_summary\n\nagent/redact.py already had the full pattern set; it was wired only to\nlog formatters and cron scrubbing, never to compression.\n\nTests: 6 regression cases covering API-key prefixes, env assignments,\nauthorization headers, JSON token fields, non-secret content\npreservation, and summarizer-echo defense.\n\nRefs: openclaw/openclaw#67801\n","name":"openclaw-port/compaction-credential-redaction","oid":"72d53e14aef8590deb7c4001449b61fc8be6b385"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T03:58:08.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T03:58:08.000000+05:30"},"message":"fix(workspace): address code review findings on parser layer\n\n- Cache MarkItDown instance across _convert() calls instead of\n recreating per file (perf during batch indexing)\n- Log warning when configured backend name is unknown in build_parser()\n- Normalize suffix casing in CompositeParser.can_parse()\n- Remove unused import in test file\n","name":"sid/workspace-salvage","oid":"cf292d258aba3428241f0f602c6eea09804d784b"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T01:39:39.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T01:39:39.000000+05:30"},"message":"ci: add path filters to Docker and test workflows, remove supply chain audit\n\n- Docker build only triggers on main push (code/config changes) and\n releases, no longer on every PR\n- Tests skip markdown-only and docs-only changes\n- Remove supply-chain-audit workflow\n","name":"chore/ci-path-filters","oid":"f7aa8b15459a5e687968fccc88535188ff1ac452"},{"author":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T01:34:30.000000+05:30"},"committer":{"email":"balyan.sid@gmail.com","name":"alt-glitch","when":"2026-04-20T01:34:30.000000+05:30"},"message":"chore(docs): remove stale documentation files\n\nRemove outdated docs that no longer reflect the current architecture:\nACP setup guide, Honcho integration spec, OpenClaw migration notes,\npricing architecture design, ink-gateway TUI migration plan,\nexample skin config, and container CLI review fixes.\n","name":"chore/remove-stale-docs","oid":"e5492dc00257db368840b0eb4cb9ea9f8150372f"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T01:01:58.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-20T01:01:58.000000+05:30"},"message":"fix(gemini): honor explicit aux endpoints and harden SSE parsing\n\n- preserve explicit Gemini base_url/api_key in auxiliary auto routing\n- make native SSE parsing handle multiline data frames correctly\n- add regression tests for auxiliary base_url precedence and SSE parsing\n- document native Gemini defaults and explicit OpenAI-compatible overrides\n","name":"feat/native-gemini-provider","oid":"fba9e024740c3ecd803f07c652741f2d40061092"},{"author":{"email":"4317663+helix4u@users.noreply.github.com","name":"helix4u","when":"2026-04-16T21:42:41.000000-06:00"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-19T20:13:59.000000+05:30"},"message":"fix(zai): autodetect endpoint during setup flow\n","name":"salvage/helix4u-zai-setup","oid":"c20577b6c0de02e82d21a4a55490fdbab7aca892"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T05:46:01.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T05:46:14.000000-07:00"},"message":"refactor(process): extend preexec_fn→start_new_session swap to remaining Popen sites\n\nPR #8399 replaced preexec_fn=os.setsid with start_new_session= in\ntools/environments/local.py to use CPython's thread-safe POSIX fastpath\ninstead of a between-fork-and-exec callback. Apply the same swap to the\nother three files covered by tests/tools/test_windows_compat.py so all\nPopen call sites use a consistent idiom.\n\nAlso:\n- Strip a trailing-whitespace line introduced by the original PR.\n- Update CONTRIBUTING.md + website/docs/developer-guide/contributing.md\n to recommend start_new_session in the cross-platform process management\n section.\n- Fix a stale docstring reference in tools/environments/base.py (_popen_bash).\n- Extend test_windows_compat.py with a parallel regression guard that\n rejects bare start_new_session=True (must be gated on _IS_WINDOWS).\n\nScope note: this is code hygiene, not a fix for #8340 — the two forms\ninvoke the same setsid() syscall, so this swap alone does not change\nbehavior for the 'setsid ... & disown' hang scenario in that issue.\n#8340 remains open.\n","name":"hermes/hermes-150d8cf8","oid":"c74030a11d531fe88a3c07d63c3113a7c2943ad7"},{"author":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T05:20:51.000000-07:00"},"committer":{"email":"teknium1@gmail.com","name":"Teknium","when":"2026-04-19T05:20:51.000000-07:00"},"message":"feat(browser): add browser_dialog for native JS dialog handling\n\nErgonomic wrapper over CDP's Page.handleJavaScriptDialog that accepts\nor dismisses alert/confirm/prompt/beforeunload dialogs blocking a page.\nUnsticks pages whose JS thread is frozen by an unhandled dialog —\nsymptom is that browser_snapshot, browser_console, browser_click etc.\nstart hanging or erroring.\n\n- action='accept'|'dismiss' required; prompt_text optional for prompt()\n- target_id auto-resolves when exactly one page tab is open; with\n multiple page tabs, errors with the tab list so the agent picks one\n- Shares browser_cdp's check_fn gate — only appears when CDP is\n reachable (/browser connect or browser.cdp_url in config). Hidden\n otherwise so backends that can't use it don't see it.\n- Safe as a probe: CDP returns a clean 'No dialog is showing' error\n when nothing's pending, which we pass through verbatim\n\nDialog detection (knowing a dialog is open without being told) is NOT\nincluded — it requires persistent CDP subscriptions per session, a\nlarger architectural change. Documented as a follow-up; agents infer\nfrom symptoms and use this tool to recover.\n\nTests: 11 new unit tests against mock CDP server covering the wrapper\n(action validation, auto-resolve with 0/1/multiple page targets,\nexplicit target_id accept/dismiss flow, prompt_text passthrough, shared\ngate with browser_cdp, registry dispatch). E2E probe case against real\nheadless Chrome passes. Positive-case real-Chrome E2E is blocked by\nChromium's headless auto-dismiss behavior when no persistent listener\nis attached — unit tests exercise the exact CDP protocol we send, so\nthe handling path is protocol-verified; headful real-browser usage\n(the actual /browser connect case) keeps dialogs alive via the Chrome\nUI.\n","name":"hermes/browser-dialog","oid":"4b8272f549c446d06395f6aec19a141b817b9810"},{"author":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-19T11:25:48.000000+05:30"},"committer":{"email":"82637225+kshitijk4poor@users.noreply.github.com","name":"kshitijk4poor","when":"2026-04-19T11:25:48.000000+05:30"},"message":"fix: token accounting fallback + reasoning-aware compression\n\nFix 1 — Token estimation fallback (closes #12023):\nWhen providers like MiniMax via OpenRouter silently ignore\nstream_options.include_usage, response.usage is None and the token\naccounting block is skipped entirely. Added an else branch that falls\nback to estimate_messages_tokens_rough() / estimate_tokens_rough() so\nsessions don't permanently record 0/0 tokens.\n\nFix 2 — Subtract reasoning tokens from compression trigger (closes #12026):\nThe compression trigger fed raw completion_tokens (including internal\nreasoning tokens) to the context compressor. For thinking models\n(GLM-5.1, QwQ, DeepSeek-R1), completion_tokens includes reasoning\nthat is NOT re-sent on subsequent turns and doesn't consume context\nwindow space. Now subtracts canonical_usage.reasoning_tokens (from\ncompletion_tokens_details.reasoning_tokens) before feeding the\ncompressor, so only content tokens count toward the threshold.\n\nThis addresses Teknium's review feedback on #12028: rather than\ndropping all completion_tokens (which would be wrong when reasoning IS\nre-sent), we use the API-provided reasoning_tokens breakdown to\nsubtract only the phantom tokens. Non-thinking models (reasoning_tokens=0)\nsee zero behavior change.\n\nProduction evidence: 6 consecutive GLM-5.1 sessions ended with\npremature compression (TD Promo #2-#6, April 17 2026). Only 3-15% of\nassistant messages had reasoning captured; total stored reasoning was\n~150-2500 tokens per session — yet completion_tokens included 15-20K\nof hidden reasoning that inflated the trigger past the 101K threshold.\n\nResearch: OpenCode has the identical bug (tui.go:335-341, completion +\nprompt without reasoning subtraction). The OpenAI/OpenRouter APIs\nprovide completion_tokens_details.reasoning_tokens for exactly this\npurpose; Hermes already extracts it via normalize_usage() but never\nused it in compression.\n\nTests: 6 new regression tests covering reasoning subtraction, premature\ncompression prevention, threshold still firing when truly full, zero\nreasoning passthrough, and fallback estimation.\n","name":"fix/minimax-glm-token-compression","oid":"eba720fc81d6366c58735de953d59a7af7b6c1d0"}],"cursor":"100","defaultBranch":"main","total":945}